ALBERT

Description: In this paper, the classification via sprepresentation and multitask learning is presented for target recognition in SAR image. To capture the characteristics of SAR image, a multidimensional generalization of the analytic signal, namely the monogenic signal, is employed. The original signal can be then orthogonally decomposed into three components: 1) local amplitude; 2) local phase; and 3) local orientation. Since the components represent the different kinds of information, it is beneficial by jointly considering them in a unifying framework. However, these components are infeasible to be directly utilized due to the high dimension and redundancy. To solve the problem, an intuitive idea is to define an augmented feature vector by concatenating the components. This strategy usually produces some information loss. To cover the shortage, this paper considers three components into different learning tasks, in which some common information can be shared. Specifically, the component-specific feature descriptor for each monogenic component is produced first. Inspired by the recent success of multitask learning, the resulting features are then fed into a joint sparse representation model to exploit the intercorrelation among multiple tasks. The inference is reached in terms of the total reconstruction error accumulated from all tasks. The novelty of this paper includes 1) the development of three component-specific feature descriptors; 2) the introduction of multitask learning into sparse representation model; 3) the numerical implementation of proposed method; and 4) extensive comparative experimental studies on MSTAR SAR dataset, including target recognition under standard operating conditions, as well as extended operating conditions, and the capability of outliers rejection.

Description: Offsets of synthetic aperture radar (SAR) images have played an important role in deriving complete three-dimensional (3-D) surface displacement fields in geoscientific applications. However, offset maps often suffer from multiple outliers and patch-like artifacts, because the standard offset-measurement method is a regular moving-window operation that does not consider the scattering characteristics of the ground. Here, we show that by focusing the offset measurements on predetected strong reflectors, the reliability and accuracy of SAR offsets can be significantly improved. Application to the 2011 Van (Turkey) earthquake reveals a clear deformation signal from an otherwise decorrelated interferogram, making derivation of the 3-D coseismic displacement field possible. Our proposed method can improve mapping of coseismic deformation and other ground displacements, such as glacier flow and landslide movement when strong reflectors exist.

Description: We present an accelerated probabilistic learning concept and its prototype implementation for mining heterogeneous Earth observation images, e.g., multispectral images, synthetic aperture radar (SAR) images, image time series, or geographical information systems (GIS) maps. The system prototype combines, at pixel level, the unsupervised clustering results of different features, extracted from heterogeneous satellite images and geographical information resources, with user-defined semantic annotations in order to calculate the posterior probabilities that allow the final probabilistic searches. The system is able to learn different semantic labels based on a newly developed Bayesian networks algorithm and allows different probabilistic retrieval methods of all semantically related images with only a few user interactions. The new algorithm reduces the computational cost, overperforming existing conventional systems, under certain conditions, by several orders of magnitude. The achieved speed-up allows the introduction of new feature models improving the learning capabilities of knowledge-driven image information mining systems and opening them to Big Data environments.

Description: The mineralogy and environmental history of Mars are being extensively studied through remote sensing observations paired with laboratory and in situ experiments. A significant portion of these experiments is being devoted to the identification and mapping of different iron oxides present in the Martian terrains. Among these compounds, goethite has been an object of great interest since its occurrence can be interpreted as mineralogical evidence of past aqueous activity on those landscapes. Although such experiments can provide valuable information regarding the presence of these minerals, the scope of the resulting observations may be hindered by logistics and cost-related constraints. We believe that predictive computer simulations can be employed to mitigate some of these constraints and contribute to the generation and validation of hypotheses in this area. Accordingly, we propose the use of SPLITS ( Sp ectral Li ght T ransport Model for S and) in investigations involving the spectral signatures of iron-bearing regions of Mars. In this paper, we initially demonstrate the predictive capabilities of the SPLITS model in this context through qualitative comparisons of modeled results with actual observations and measured data. Using the resulting modeled reflectance curves as our baseline data, we then perform a series of controlled computational experiments to investigate how variations on goethite and hematite content affect the spectral responses of Martian sand-textured soils.

Description: Thermal infrared (TIR) time series are exploited by many methods based on Earth observation (EO), for such applications as agriculture, forest management, and meteorology. However, due to physical limitations, data acquired by a single sensor are often unsatisfactory in terms of spatial or temporal resolution. This issue can be tackled by using remotely sensed data acquired by multiple sensors with complementary features. When nonreal-time functioning or at least near real-time functioning is admitted, the measurements can be profitably fed to a sequential Bayesian algorithm, which allows to account for the correlation embedded in the successive acquisitions. In this work, we focus on applications that allow the batch processing of the whole data sequences acquired in a fixed time interval. In this case, multiple options for improving the final product are offered by the Bayesian framework, based on both sequential and smoothing techniques. We consider several such Bayesian strategies and comparatively assess their performances in practical applications and through real thermal data acquired by the SEVIRI and MODIS sensors, encompassing the presence of multiple disturbance source, e.g., the cloud coverage of the illuminated scene.

Description: Land surface albedo, qualifying the ratio of the radiant flux reflected from the land surface to the incident flux, is a key forcing parameter controlling the Earth’s energy budget. Previously, several BRDF archetypes were distilled from high-quality MODIS BRDF/Albedo products. In this study, we propose a method that largely relies on matching observed multiangular reflectances with the most appropriate of these prior BRDF archetypes to determine the amplitude and shape of the actual surface BRDFs, when directional signatures are insufficient. This method is first evaluated using an assortment of multisource BRDF data sets to demonstrate its viability for surface albedo estimates, and then is applied to airborne wide-angle infrared dual-mode line/area array scanner (WIDAS) from the Watershed Allied Telemetry Experimental Research (WATER) campaign in the Heihe River Basin of China in 2008. This algorithm makes use of the linear MODIS BRDF model to determine the BRDF archetypes needed as prior knowledge for intrinsic spectral albedo estimates. The intrinsic spectral albedos are then used to estimate actual spectral albedos by considering the proportion of direct and diffuse solar radiation. A spectral-to-broadband conversion is performed to generate the broadband albedo at shortwave regimes through the use of conversion coefficients derived from extensive radiative transfer simulations. A further validation confirms that the estimated albedos are consistent with in situ field measured albedos over available corn crop sites. This method provides a major advantage on utilizing generalized BRDF information derived from MODIS in conjunction with other instrument data that are acquired with less angular variation.

Description: Ocean current is highly related to the interaction between ocean and atmosphere. By measuring the speed and direction of the ocean current from space, we can investigate the ocean–atmosphere interaction on a global scale. The ocean–atmosphere interaction helps to maintain the balance that is essential for planet habitability. However, the conventional scatterometer is unable to measure the ocean current vector. To achieve this, a potentially feasible approach is to use a bigger antenna, a higher PRF, and measure the interferometric phase of two successive echoes. This paper derives four decorrelation factors, and provides the phase error model first. Then, an end-to-end simulation model is established, and it is used to analyze the feasibility of ocean surface current measurement from space. Based on the simulation model, the system parameters are optimized. The simulation results show that the current speed standard deviation (Std), which means the measurement accuracy, in along-track and cross-track direction is smaller than 0.1 m/s when the wind speed is larger than 4 m/s. The swath can be used for current vector inversion that is greater than 70% when the wind speed is larger than 7 m/s. Meanwhile, ${{bf K}_{{bf pc}}}$ of the modified scatterometer is computed and the results show that ${{bf K}_{{bf pc}}}$ is better than the traditional pencil-beam rotating scatterometer when the wind speed is larger than 6 m/s.

Description: High-resolution three-dimensional (3-D) radar imaging of space targets in micromotion plays a significant role in space target recognition and space situation awareness; thus, it has attracted extensive attention in recent years. Because of the fast rotation, some scattering centers are occluded by others, i.e., the scattering centers cannot be continuously illuminated by radar in the imaging interval, and their radar echoes are discontinuous. In this paper, a nonparametric 3-D imaging method based on scattering center trajectory association is proposed. It deals with target occlusion using the Riemannian manifold optimization and obtains focused imaging of targets in complex micromotion. The effectiveness of the proposed method is validated using simulated data.

Description: Recently, satellite-based systems have been introduced that utilize angle-of-arrival (AOA) measurements to geo-locate objects of interest. In the previous work, we considered the application of nonlinear optimization to AoA-based geolocation to these systems. This previous work, however, assumed that all noise sources were independent. In the case of fast-moving objects, however, there is a significant source of error due to the propagation time inherent in satellite-based observation of objects due to the difference between the location of the object when it is observed by a satellite, and the location of the object when it emitted the signal that is being measured. This introduces a systematic error into the system that cannot be resolved by the system proposed by Burchett et al. In this paper, we extend our prior work to account for the time-delay inherent in satellite-based geolocation systems, making this system accurate for fast-movers as well as fixed or slow-moving objects. Results demonstrating significant improvement in geolocation performance both in terms of accuracy and estimated error bounds are presented.

Description: Most of existing change detection methods could be classified into three groups, the traditional pixel-based change detection (PBCD), the object-based change detection (OBCD), and the hybrid change detection (HCD). Nevertheless, both PBCD and OBCD have disadvantages, and classical HCD methods belong to intuitive decision-level fusion schemes of PBCD and OBCD. There is no optimum HCD method as of yet. Analyzing the complementarities of PBCD and OBCD method, we propose a new unsupervised algorithm-level fusion scheme (UAFS-HCD) in this paper to improve the accuracy of PBCD using spatial context information through: 1) getting the preliminary change mask with PBCD at first to estimate some parameters for OBCD; 2) deriving the unchanged area mask to eliminate the areas without changes, reducing error amplification phenomenon of OBCD; and 3) obtaining the final change mask by means of OBCD method. Taking flood detection with multitemporal SAR data as an example, we compared the new scheme with some classical methods, including PBCD, OBCD, and HCD method and supervised manual trial-and-error procedure (MTEP). The experimental results of flood detection showed that the new scheme was efficient and robust, and its accuracy sometimes can even exceed MTEP.

Description: This paper describes a methodology to extract a consistent human settlement extent layer using Landsat data and its implementation in the Google Earth Engine platform. The approach allows the extraction of human settlement extents by means of the existing Landsat 5 and 7 data sets, allowing to check their evolution at 30-m spatial resolution. Since human settlements are the main proxy to people geographical distribution and to building locations, this layer may serve as a mean to disaggregate people/building counts at the regional/national level. The approach is tested in several parts of the world against existing ground truth data at the same spatial resolution in Brazil and China, as well as against extents manually extracted from VHR data in three different geographical areas: 1) Brazil; 2) South East China; and 3) Indonesia.

Description: Remote sensing of urban areas is mainly conducted at high frequencies, to obtain highly resolved images for classification, target detection, or urban areas monitoring for instance. We propose on the contrary to investigate the use of VHF-band for observation, between 120 and 360 MHz. In particular, our concern is to allow target detection, when the object is not in the line of sight (LOS) of the radar, which is typically an issue in urban areas. The benefits of low frequencies are highligted using measurements and simulations over a simple metallic model of two buildings surrounding a street.

Description: The atmospheric condition parameters used in the radiative transfer-based atmospheric correction (AC) are often uncertain. This uncertainty propagates to the estimated reflectance. The reflectance, is, however, not equally sensitive to all the parameters. A sensitivity analysis (SA) helps in prioritizing the parameters. The objective of this study was to perform an SA of reflectance to water vapor concentration ( $wv$ ) and aerosol optical thickness ( $AOT$ ). SA was performed using the Fourier amplitude sensitivity test (FAST) method, which computes sensitivity indices ( $text{SI}$ ) of these parameters. Besides variation in the two parameters, we also studied the effect of surface albedo on the $text{SI}$ by quantifying $text{SI}$ for three target surfaces (in the spectral range $text{0.44{-}0.96};upmu$ m): 1) a dark target (water); 2) a bright target (bare soil); and 3) a target having low albedo in the visible and high albedo in near-infrared range (forest). For $AOT$ , high ( $approx!text{0.9}$ ) $text{SI}$ values were observed at the nonwater absorption wavelengths. For $wv$ , high $text{SI}$ values were observed at wavelengths, where strong absorption features are loca- ed and when the surface albedo was high. For the dark target, the effect of $AOT$ was prominent throughout the spectral range. We found that the sensitivity of reflectance to $wv$ and $AOT$ is a function of wavelength, strength of the absorption features, and surface albedo. We conclude that $AOT$ is a more important parameter for dark targets than $wv$ even at the principal absorption feature. For bright targets, the importance of $wv$ and $AOT$ depends on the strength of the absorption feature.

Description: The feasibility study of the HALESIS (High-Altitude Luminous Events Studied by Infrared Spectro-imagery) project is presented. The purpose of this experiment is to measure the atmospheric perturbation in the minutes following the occurrence of transient luminous events (TLEs) from a stratospheric balloon in the altitude range of 20–40 km. The instrumentation will include a spectro-imager embedded in a pointing gondola. Infrared signatures of a single blue jet were simulated under the assumption of local thermodynamic equilibrium (LTE), and were then compared with a panel of commercially available instrument specifications. The sensitivity of the signatures with a local perturbation of the main vibrational energy level populations of ${mathbf{rm CO}_{2}}$ , CO, NO, ${mathbf{rm O}_{3}}$ , and ${mathbf{rm H}_{2}}{rm O}$ was measured and the infrared signatures of a single blue jet taking into account non-LTE hypotheses were compared with the same panel of commercially available instrument specifications. Lastly, the feasibility of the study is discussed.

Description: To improve accuracy and efficiency of object detection and classification with hyperspectral imagery (HSI), we propose a novel smoothing algorithm by coupling of a Laplacian-based reaction term to a classical divergence-based anisotropic diffusion partial differential equation (PDE). In addition, an adaptive parameter is introduced to regularize this nonlinear reaction-diffusion PDE by explicitly integrating the interband correlations with the noise level of each band in HSI. It is also well-known that the interband correlations can be implicitly embedded into the diffusion coefficient of the divergence-based PDE, to allow a selective smoothing that reduces the local homogeneous area variability while preventing smoothing across edges. Therefore, the interband correlations in HSI are exploited in the proposed method in both explicit and implicit ways. As a result, our algorithm is more effective at controlling the behavior of the diffusion evolution when compared to previous multi/hyperspectral diffusion algorithms. The simulations based on both synthetic data and real hyperspectral remote sensing data show that our algorithm can improve the hyperspectral data quality in terms of both visual inspection and image quality indices.

Description: This paper proposes a novel approach to classify hyperspectral (HS) images using both spectral and spatial information. It first consists of a supervised spectral dimension reduction step that transforms the HS image into a score image that has fewer channels. These channels are chosen so as to enhance distances between classes to be discriminated and to reduce background variability, thus leading to edges that correspond to actual class borders. In the second step, applying an edge-preserving spatial regularization on this score image leads to a lowered background variability. Therefore, in the third step, the pixel-wise classification of the regularized score image is greatly improved. We implement this approach using the partial least squares (PLS) method for spectral dimension reduction and the anisotropic diffusion for spatial regularization. We then compare linear discriminant analysis (LDA), K-nearest neighbors (KNN), and support vector machine (SVM) for the class decision. The effectiveness of our method was evaluated with three remotely sensed HS images. Its robustness was also assessed for different training sets, since the latter has a crucial influence on classification performance. On average, our method gave better results in terms of classification accuracy and was more robust than other classification methods tested with the same images.

Description: Due to its simple, fast, and good generalization ability, extreme learning machine (ELM) has recently drawn increasing attention in the pattern recognition and machine learning fields. To investigate the performance of ELM on the hyperspectral images (HSIs), this paper proposes two spatial–spectral composite kernel (CK) ELM classification methods. In the proposed CK framework, the single spatial or spectral kernel consists of activation–function-based kernel and general Gaussian kernel, respectively. The proposed methods inherit the advantages of ELM and have an analytic solution to directly implement the multiclass classification. Experimental results on three benchmark hyperspectral datasets demonstrate that the proposed ELM with CK methods outperform the general ELM, SVM, and SVM with CK methods.

Description: This paper proposes a framework for hyperspectral images (HSIs) classification with composite kernels discriminant analysis (CKDA). The CKDA uses the spectral and spatial information extracted by Gaussian weighted local mean operator (GWLM) and is suitable to solve few labeled samples classification problem of HSI, which has very important practical significance for the case that training samples are insufficient due to high cost. Experimental results show that the spatial information extracted by GWLM can greatly improve the performance, and demonstrate the superiority of CKDA for HSI classification in the case of few labeled samples. Compared with other state-of-the-art spectral–spatial kernel methods, the proposed methods also show very good advantages, especially the parallel kernel method.

Description: Accurate generation of a land cover map using hyperspectral data is an important application of remote sensing. Multiple classifier system (MCS) is an effective tool for hyperspectral image classification. However, most of the research in MCS addressed the problem of classifier combination, while the potential of selecting classifiers dynamically is least explored for hyperspectral image classification. The goal of this paper is to assess the potential of dynamic classifier selection/dynamic ensemble selection (DCS/DES) for classification of hyperspectral images, which consists in selecting the best (subset of) optimal classifier(s) relative to each input pixel by exploiting the local information content of the image pixel. In order to have an accurate as well as computationally fast DCS/DES, we proposed a new DCS/DES framework based on extreme learning machine (ELM) regression and a new spectral–spatial classification model, which incorporates the spatial contextual information by using the Markov random field (MRF) with the proposed DES method. The proposed classification framework can be considered as a unified model to exploit the full spectral and spatial information. Classification experiments carried out on two different airborne hyperspectral images demonstrate that the proposed method yields a significant increase in the accuracy when compared to the state-of-the-art approaches.

Description: Advanced classifiers, e.g., partial least squares discriminant analysis (PLS-DA) and random forests (RF), have been recently used to model reflectance spectral data in general, and of soil properties in particular, since their spectra are multivariate and highly collinear. Preprocessing transformations (PPTs) can improve the classification accuracy by increasing the variability between classes while decreasing the variability within classes. Such PPTs are common practice prior to a PLS-DA, but are rarely used for RF. The objectives of this paper are twofold: to compare the performances of PLS-DA and RF for modeling the spectral reflectance of soil in changed land-uses with different treatments and to compare the effects of nine different PPTs on the prediction accuracy of each of these classification methods. Differences in six physical, biological, and chemical soil properties of changed land-uses from the northern Negev Desert in Israel were evaluated. Significant differences were found between soil properties, which are used to classify land-uses and treatments. Depending on the dataset, different PPTs improved the classification accuracy by 11%–24% and 32%–42% for PLS-DA and RF, respectively, in comparison to the spectra without PPT. Out of the PPTs tested, the generalized least squares weighting (GLSW)-based transformations were found to be the most effective for most classifications using both PLS-DA and RF. Our results show that both PLS-DA and RF are suitable classifiers for spectral data, provided that an appropriate PPT is applied.

Description: This paper contributes the concept of spectral–spatial kernel-based multivariate analysis (KMVSSA) based on the statistical principle of multivariate statistics. The essence of proposed framework is to expose the inherent structure and meaning revealed within spectral and spatial features through various statistical methods in hyperspectral remotely sensed data. This kernel-based framework is investigated to incorporate the spectral and spatial information simultaneously for dimension reduction and classification of hyperdimensional datasets. The method uses multivariate analysis to choose and apply a transform matrix that the transformed components are as orthogonal as possible. This nonlinear framework is derived by means of the theory of complete orthonormal systems. KMVSSA exhibits great flexibility by the combination of spectral and spatial features. We investigate the possibility of using KMVSSA for the classification of hyperspectral images and dimension reduction. The proposed framework is examined and compared in different merits with several hyperspectral images in different conditions (urban/agricultural area and size of the training set). Experimental results show that the proposed framework can meaningfully enhance the dimensionality reduction and also it greatly improves the overall as well as per class classification accuracies. We demonstrate a comprehensive comparison of some state of the art hyperspectral image classification methods.

Description: Development of intelligent decision-making systems for complex problems, such as land covers classification of hyperspectral remote sensing (HSRS) images, requires efficient interpretation of available information through conceptual rather than numerical level. Granular neural network (GNN) in combination with the granular representation of information using linguistic terms is one such system. GNN takes the fuzzified input information and processes them with neural network (NN) architecture, where the network structure is transparent enough to interpret the processing steps. Further, knowledge encoding has been considered as one of the principal elements of intelligent decision-making systems. This paper proposes a new model of knowledge-encoded GNNs for land cover classification of HSRS images. Knowledge encoding is done using neighborhood rough sets (NRSs) that explore the local/contextual information. The encoded knowledge using NRS is obtained in the form of dependency rules with respect to the output class labels of land covers and these rules determine appropriate number of hidden nodes of GNNs. The dependency factors obtained during rule generation are used for initializing the connecting weights of GNNs. NRS is also used here in the selection of a subset of features for reducing the burden of high-dimensional fuzzy-granulated feature space of HSRS image. The proposed model thus explores jointly the advantages of fuzzy granulation, GNNs, and feature selection and knowledge encoding using NRS. Superiority of the model to similar other methods are justified in land covers classification of two HSRS images acquired by different remotely placed sensors.

Description: Several popular endmember extraction and unmixing algorithms are based on the geometrical interpretation of the linear mixing model, and assume the presence of pure pixels in the data. These endmembers can be identified by maximizing a simplex volume, or finding maximal distances in subsequent subspace projections, while unmixing can be considered a simplex projection problem. Since many of these algorithms can be written in terms of distance geometry, where mutual distances are the properties of interest instead of Euclidean coordinates, one can design an unmixing chain where other distance metrics are used. Many preprocessing steps such as (nonlinear) dimensionality reduction or data whitening, and several nonlinear unmixing models such as the Hapke and bilinear models, can be considered as transformations to a different data space, with a corresponding metric. In this paper, we show how one can use different metrics in geometry-based endmember extraction and unmixing algorithms, and demonstrate the results for some well-known metrics, such as the Mahalanobis distance, the Hapke model for intimate mixing, the polynomial post-nonlinear model, and graph-geodesic distances. This offers a flexible processing chain, where many models and preprocessing steps can be transparently incorporated through the use of the proper distance function.

Description: Progress in mapping plant species remotely with imaging spectroscopy data is limited by the traditional classification framework, which carries the requirement of exhaustively defining all classes (species) encountered in a landscape. As the research objective may be to map only one or a few species of interest, we need to explore alternative classification methods that may be used to more efficiently detect a single species. We compared the performance of three support vector machine (SVM) methods designed for single-class detection—binary (one-against-all) SVM, one-class SVM, and biased SVM—in detecting five focal tree and shrub species using data collected by the Carnegie Airborne Observatory over an African savanna. Prior to this comparison, we investigated the effects of training data amount and balance on binary SVM and evaluated alternative methods for tuning one-class and biased SVMs. A key finding was that biased SVM was generally best parameterized by crown-level cross validation paired with the tuning criterion proposed by Lee and Liu [1] . Among the different single-class methods, binary SVM showed the best overall performance (average F-scores 0.43–0.78 among species), whereas one-class SVM showed very poor performance (F-scores 0.09–0.46). However, biased SVM produced results similar to those obtained with binary SVM (F-scores 0.40–0.72), despite using labeled training data from only the focal class. Our results indicate that both binary and biased SVMs can work well for remote single-species detection, while both methods, particularly biased SVM, greatly reduce the amount of training data required compared with traditional multispecies classification.

Description: Hundreds of narrow contiguous spectral bands collected by a hyperspectral sensor have provided the opportunity to identify the various materials present on the surface. Moreover, spatial information, enforcing the assumption that the adjacent pixels belong to the same class with a high probability, is a valuable complement to the spectral information. In this paper, two predominant approaches have been developed to exploit the spatial information. First, by decomposing each pixel and the spatial neighborhood into a low-rank form, the spatial information can be efficiently integrated into the spectral signatures. Meanwhile, in order to describe the low-rank structure of the decomposed data more precisely, an $ell_{1/2}$ norm regularization is introduced and a discrete algorithm is proposed to solve the combined optimization problem by the augmented Lagrange multiplier (ALM) and a half-threshold operator. Second, a graph cuts segmentation algorithm has been applied on the sparse-representation-based probability estimates of the hyperspectral data to further improve the spatial homogeneity of the material distribution. Experimental results on four real hyperspectral data with different spectral and spatial resolutions have demonstrated the effectiveness and versatility of the proposed spatial information-fused approaches for hyperspectral image classification.

Description: Classification of hyperspectral imagery using few labeled samples is a challenging problem, considering the high dimensionality of hyperspectral imagery. Classifiers trained on limited samples with abundant spectral bands tend to overfit, leading to weak generalization capability. To address this problem, we have developed an enhanced ensemble method called multiclass boosted rotation forest (MBRF), which combines the rotation forest algorithm and a multiclass AdaBoost algorithm. The benefit of this combination can be explained by bias-variance analysis, especially in the situation of inadequate training samples and high dimensionality. Furthermore, MBRF innately produces posterior probabilities inherited from AdaBoost, which are served as the unary potentials of the conditional random field (CRF) model to incorporate spatial context information. Experimental results show that the classification accuracy by MBRF as well as its integration with CRF consistently outperforms the other referenced state-of-the-art classification methods when limited labeled samples are available for training.

Description: Curse of dimensionality is a major disadvantage for classification of hyperspectral imagery since a large number of bands need to be dealt with. Band selection is a task to reduce the number of bands. An unsupervised band selection method is proposed in this article. It is a three-step procedure. In the first step, characteristics (attributes) of the bands are found out. Next, redundancy among the bands is removed by executing clustering operation. At last, the remaining bands, which are nonredundant among themselves, are ranked according to their discriminating capability. Discriminating capability is calculated by measuring the capacitory discrimination of the bands. Results are compared with four state-of-the-art methods: a band elimination method, a ranking-based, and two clustering-based band selection methods to demonstrate the effectiveness of the proposed method. Four evaluation measures, namely: 1) classification accuracy; 2)  Kappa coefficient; 3) class separability, and 4) entropy, are calculated over the selected bands to assess the efficiency of the selected bands. The proposed method shows promising results compared to them.

Description: A fast forward feature selection algorithm is presented in this paper. It is based on a Gaussian mixture model (GMM) classifier. GMM are used for classifying hyperspectral images. The algorithm selects iteratively spectral features that maximizes an estimation of the classification rate. The estimation is done using the k-fold cross validation (k-CV). In order to perform fast in terms of computing time, an efficient implementation is proposed. First, the GMM can be updated when the estimation of the classification rate is computed, rather than re-estimate the full model. Secondly, using marginalization of the GMM, submodels can be directly obtained from the full model learned with all the spectral features. Experimental results for two real hyperspectral data sets show that the method performs very well in terms of classification accuracy and processing time. Furthermore, the extracted model contains very few spectral channels.

Description: Band selection is an essential step toward effective and efficient hyperspectral image classification. Traditional supervised band selection methods are often hindered by the problem of lacking enough training samples. To address this problem, we propose a semisupervised band selection method that allows contribution from both labeled and unlabeled hyperspectral pixels. This method first builds a hypergraph model from all hyperspectral samples to measure the similarity among pixels. We show that hypergraph can capture relationship among pixels in both spectral and spatial domain. In the second step, a semisupervised learning method is introduced to propagate class labels to unlabeled samples. Then a linear regression model with group sparsity constraint is used for band selection. Finally, hyperspectral pixels with selected bands are used to train a support vector machine (SVM) classifier. The proposed method is tested on three benchmark datasets. Experimental results demonstrate its advantages over several other band selection methods.

Description: Spectral unmixing is an important task for remotely sensed hyperspectral data exploitation. It amounts the identification of pure spectral signatures ( endmembers ) in the data, and the estimation of the abundance of each endmember in each (possibly mixed) pixel. A challenging problem in spectral unmixing is how to determine the number of endmembers in a given scene. One of the most popular and widely used techniques for this purpose is the HySime algorithm but, due to the complexity and high dimensionality of hyperspectral scenes, this technique is computational expensive. Reconfigurable field-programmable gate arrays (FPGAs) are promising platforms that allow hardware/software codesign and the potential to provide powerful onboard computing capabilities and flexibility at the same time. In this paper, we present the first FPGA design for the HySime algorithm. Our system includes a direct memory access (DMA) module and implements a prefetching technique to hide the latency of the input/output communications. The proposed method has been implemented on a Virtex-7 XC7VX690T FPGA and tested using real hyperspectral data collected by NASAs airborne visible infra-red imaging spectrometer (AVIRIS) over the Cuprite mining district in Nevada and the World trade center (WTC) in New York. Experimental results demonstrate that our hardware version of the HySime algorithm can significantly outperform a software version, which makes our reconfigurable system appealing for onboard hyperspectral data processing.

Description: Spatial–spectral classification is a very important topic in the field of remotely sensed hyperspectral imaging. In this work, we develop a parallel implementation of a novel supervised spectral–spatial classifier, which models the likelihood probability via ${bm{l}_{mathbf{1}}} - {bm{l}_{mathbf{2}}}$ sparse representation and the spatial prior as a Gibbs distribution. This classifier takes advantage of the spatial piecewise smoothness and correlation of neighboring pixels in the spatial domain, but its computational complexity is very high which makes its application to time-critical scenarios quite limited. In order to improve the computational efficiency of the algorithm, we optimized its serial version and developed a parallel implementation for commodity graphics processing units (GPUs). Our parallel spatial–spectral classifier with sparse representation and Markov random fields (SSC-SRMRF-P) exploits the low-level architecture of GPUs. The parallel optimization of the proposed method has been carried out using the compute unified device architecture (CUDA). The performance of the parallel implementation is evaluated and compared with the serial and multicore implementations on central processing units (CPUs). In fact, the proposed method has been designed to adequately exploit the massive data parallel capacities of GPUs together with the control and logic capacities of CPUs, thus resorting to a heterogeneous CPU–GPU framework in the design of the parallel algorithm. Experimental results using real hyperpsectral images demonstrate very high performance for the proposed CPU–GPU parallel method, both in terms of classification accuracy and computational performance.

Description: Classification is one of the most important analysis techniques for hyperspectral image analysis. Sparse representation is an extremely powerful tool for this purpose, but the high computational complexity of sparse representation-based classification techniques limits their application in time-critical scenarios. To improve the efficiency and performance of sparse representation classification techniques for hyperspectral image analysis, this paper develops a new parallel implementation on graphics processing units (GPUs). First, an optimized sparse representation model based on spatial correlation regularization and a spectral fidelity term is introduced. Then, we use this approach as a case study to illustrate the advantages and potential challenges of applying GPU parallel optimization principles to the considered problem. The first GPU optimization algorithm for sparse representation classification (SRCSC_P) of hyperspectral images is proposed in this paper, and a parallel implementation of the proposed method is developed using compute unified device architecture (CUDA) on GPUs. The GPU parallel implementation is compared with the serial and multicore implementations on CPUs. Experimental results based on real hyperspectral datasets show that the average speedup of SRCSC_P is more than $mathbf{130} times$ , and the proposed approach is able to provide results accurately and fast, which is appealing for computationally efficient hyperspectral data processing.

Description: This paper reports the outcomes of the 2014 Data Fusion Contest organized by the Image Analysis and Data Fusion Technical Committee (IADF TC) of the IEEE Geoscience and Remote Sensing Society (IEEE GRSS). As for previous years, the IADF TC organized a data fusion contest aiming at fostering new ideas and solutions for multisource remote sensing studies. In the 2014 edition, participants considered multiresolution and multisensor fusion between optical data acquired at 20-cm resolution and long-wave (thermal) infrared hyperspectral data at 1-m resolution. The Contest was proposed as a double-track competition: one aiming at accurate landcover classification and the other seeking innovation in the fusion of thermal hyperspectral and color data. In this paper, the results obtained by the winners of both tracks are presented and discussed.

Description: As current and future satellite systems provide both hyperspectral and multispectral images, a need has arisen for image fusion using hyperspectral and multispectral images to improve the fusion quality. This study introduces a hyperspectral image fusion algorithm using multispectral images with a higher spatial resolution and partially different wavelength range compared with the corresponding hyperspectral images. This study focuses on an image fusion technique that enhances the spatial quality and preserves the spectral information of hyperspectral images. The proposed algorithm generates a simulated multispectral band via a spectral unmixing technique and extracts high-frequency information based on blocks of associated bands. The algorithm was applied to Compact Airborne Spectrographic Imager (CASI) datasets acquired in two modes and was compared with two existing methods. Although the wavelength range of the multispectral image did not coincide with that of the hyperspectral image, the proposed algorithm efficiently improved the spatial details and preserved the spectral information of the fused results.

Description: Restoration is important in preprocessing hyperspectral images (HSI) to improve their visual quality and the accuracy in target detection or classification. In this paper, we propose a new low-rank spectral nonlocal approach (LRSNL) to the simultaneous removal of a mixture of different types of noises, such as Gaussian noises, salt and pepper impulse noises, and fixed-pattern noises including stripes and dead pixel lines. The low-rank (LR) property is exploited to obtain precleaned patches, which can then be better clustered in our spectral nonlocal method (SNL). The SNL method takes both spectral and spatial information into consideration to remove mixed noises as well as preserve the fine structures of images. Experiments on both synthetic and real data demonstrate that LRSNL, although simple, is an effective approach to the restoration of HSI.

Description: Extreme learning machine (ELM) is an efficient learning algorithm that has been recently applied to hyperspectral image classification. In this paper, the first implementation of the ELM algorithm fully developed for graphical processing unit (GPU) is presented. ELM can be expressed in terms of matrix operations so as to take advantage of the single instruction multiple data (SIMD) computing paradigm of the GPU architecture. Additionally, several techniques like the use of ensembles, a spatial regularization algorithm, and a spectral–spatial classification scheme are applied and projected to GPU in order to improve the accuracy results of the ELM classifier. In the last case, the spatial processing is based on the segmentation of the hyperspectral image through a watershed transform. The experiments are performed on remote sensing data for land cover applications achieving competitive accuracy results compared to analogous support vector machine (SVM) strategies with significantly lower execution times. The best accuracy results are obtained with the spectral–spatial scheme based on applying watershed and a spatially regularized ELM.

Description: Leaf area index (LAI) is a basic quantity indicating crop growth situation and plays a significant role in ecological model and interaction model between earth surface and atmosphere. However, nonlinear estimation processes of LAI from heterogeneous remote sensing data would induce a scaling bias. The purpose of this study is to provide a method to evaluate and correct the scaling bias. For the effectiveness of the method, first both statistical and physical models were built to estimate LAI directly from modified soil-adjusted vegetation index (MSAVI) as a function with univariate and also from red and near infrared reflectances as a bivariate model. The analysis of wavelet transform and fractal theory revealed that the scaling bias and the high-frequency coefficient from LAI at fine resolution decomposed by Haar wavelet were fractal relation. Based on the wavelet–fractal method, scaling bias could be well denoted by high-frequency coefficient in log–log coordinate for both univariate model and bivariate model, and the root-mean-square error (RMSE) and relative error (RE) of estimated LAI caused by the scaling bias could be greatly reduced after scaling correction. Additionally, to analyze the influence of spatial heterogeneity and nonlinearity of the retrieval model, the scaling bias was investigated on horizontal comparison of LAI retrieval models with univariate and bivariate at a certain resolution and was longitudinally discussed in a retrieval model at different aggregation scales. This study suggests that it is feasible to successfully correct and analyze the scaling bias using the wavelet–fractal method.

Description: Hyperspectral unmixing is an important technique for remotely sensed hyperspectral data exploitation. Linear spectral unmixing is frequently used to characterize mixed pixels in hyperspectral data. Over the last few years, many techniques have been proposed for identifying pure spectral signatures (endmembers) in hyperspectral images. The iterated constrained endmembers (ICE) algorithm is an iterative method that uses the linear model to extract endmembers and abundances simultaneously from the data set. This approach does not necessarily require the presence of pixels in the hyperspectral image as it can automatically derive the signatures of endmembers even if these signatures are not present in the data. As it is the case with other endmember identification algorithms, ICE suffers from high computational complexity. In this paper, a complete and scalable adaptation of the ICE algorithm is implemented using the parallel nature of commodity graphics processing units (GPUs). This gives significant speed increase over the traditional ICE method and allows for processing of larger data set with an increased number of endmembers.

Description: Detecting and mapping plant invaders using hyperspectral remote sensing is necessary in mitigating the extensive ecologic and economic damage these alien plants induce on our forest ecosystems. Using AISA Eagle image data, this study investigated the capability of two unsupervised classification methods for the detection and mapping of Solanum mauritianum located within commercial forestry ecosystems. The existing random forest (RF) outlier detection method when used in conjunction with Anselins Moran’s I produced a detection rate (DR) of 89% with a false positive rate (FPR) of 9.26%. In comparison, the newly developed methodology which is based on the decomposition of the RF proximity matrix using principal component analysis (PCA) resulted in a DR of 95% with a lower FPR (6.39%). Overall, this research has demonstrated the potential of utilizing an unsupervised and accurate RF framework for the detection and mapping of alien invasive plants.

Description: Hyperspectral image classification has been an active topic of research. In recent years, it has been found that light detection and ranging (LiDAR) data provide a source of complementary information that can greatly assist in the classification of hyperspectral data, in particular when it is difficult to separate complex classes. This is because, in addition to the spatial and the spectral information provided by hyperspectral data, LiDAR can provide very valuable information about the height of the surveyed area that can help with the discrimination of classes and their separability. In the past, several efforts have been investigated for fusion of hyperspectral and LiDAR data, with some efforts driven by the morphological information that can be derived from both data sources. However, a main challenge for the learning approaches is how to exploit the information coming from multiple features. Specifically, it has been found that simple concatenation or stacking of features such as morphological attribute profiles (APs) may contain redundant information. In addition, a significant increase in the number of features may lead to very high-dimensional input features. This is in contrast with the limited number of training samples often available in remote-sensing applications, which may lead to the Hughes effect. In this work, we develop a new efficient strategy for fusion and classification of hyperspectral and LiDAR data. Our approach has been designed to integrate multiple types of features extracted from these data. An important characteristic of the presented approach is that it does not require any regularization parameters, so that different types of features can be efficiently exploited and integrated in a collaborative and flexible way. Our experimental results, conducted using a hyperspectral image and a LiDAR-derived digital surface model (DSM) collected over the University of Houston campus and the neighboring urban area, indicate that the proposed fram- work for multiple feature learning provides state-of-the-art classification results.

Description: The availability of graphics processing units (GPUs) provides a low-cost solution to real-time processing, which may benefit many remote sensing applications. In this paper, a spectral–spatial classification scheme for hyperspectral images is specifically adapted for computing on GPUs. It is based on wavelets, extended morphological profiles (EMPs), and support vector machine (SVM). Additionally, a preprocessing stage is used to remove noise in the original hyperspectral image. The local computation of the techniques used in the proposed scheme makes them particularly suitable for parallel processing by blocks of threads in the GPU. Moreover, a block-asynchronous updating process is applied to the EMP to speedup the morphological reconstruction. The results over different hyperspectral images show that the execution can be speeded up to $8.2times$ compared to an efficient OpenMP parallel implementation, achieving real-time hyperspectral image classification while maintaining the high classification accuracy values of the original classification scheme.

Description: This study is an application of the hybrid regression method to improve the accuracy of retrieved humidity profile from infrared hyperspectral sounding observations. In hybrid regression method, a weighted average of two regression products that are derived using two different forms of predictand is computed. Regression coefficients for each form of predictand are computed using principal components of MetOp-IASI radiance spectra. First regression product uses logarithm of specific humidity as predictand, whereas second regression product uses only specific humidity as predictand. The weights used in hybrid regression are computed at different pressure levels based on error statistics of humidity retrieval from different predictands. The hybrid regression-based method shows improvement over the state-of-the-art regression method. Humidity profiles retrieved from different regression methods are validated with collocated ECMWF humidity profiles and radiosonde observations for dry, wet, and combined atmospheric conditions. For all cases, humidity retrieved from hybrid regression method is found to be the most accurate at all pressure levels.

Description: Target and anomaly detection are important techniques for remotely sensed hyperspectral data interpretation. Due to the high dimensionality of hyperspectral data and the large computational complexity associated to processing algorithms, developing fast techniques for target and anomaly detection has received considerable attention in recent years. Although several high-performance architectures have been evaluated for this purpose, field programmable gate arrays (FPGAs) offer the possibility of onboard hyperspectral data processing with low-power consumption, reconfigurability and radiation tolerance, which make FPGAs a relevant platform for hyperspectral processing. In this paper, we develop a novel FPGA-based technique for efficient target detection in hyperspectral images. The proposed method uses a streaming background statistics (SBS) approach for optimizing the constrained energy minimization (CEM) and Reed-Xiaoli (RX) algorithms, which are widely used techniques for target and anomaly detection, respectively. Specifically, these two algorithms are implemented in streaming fashion on FPGAs. Most importantly, we present a dual mode that implements a flexible datapath to decide in real time which one among these two algorithms should be used, thus allowing for the dynamic adaptation of the hardware to either target detection or anomaly detection scenarios. Our experiments, conducted with several well-known hyperspectral scenes, indicate the effectiveness of the proposed implementations.

Description: A sensor aboard an unmanned aerial vehicle (UAV) is vulnerable to vibration and natural conditions such as erratic winds. Considering the difference between laboratory and vicarious environments of calibration, a vicarious calibration is closer to the real environment and is a complement to laboratory calibrations for remote sensors. The existing vicarious calibration of UAVs only uses a reflectance-based method, rather than irradiance-based method. Therefore, the error caused by aerosol-type assumptions, which is the largest uncertainty for reflectance-based method, is not considered sufficiently during vicarious calibration of UAVs. Considering the difference in the upward radiative transfer path between satellites and UAVs, we propose an improved irradiance-based method. A simulation experiment was designed to compare the relative differences between two aerosol types under different aerosol optical thicknesses (AOTs) and heights for both the reflectance-based and the improved methods. Additionally, two field-calibration campaigns, under different weather conditions, were performed to calibrate a Headwall hyperspectral imager payload on a UAV, with the help of calibration tarps and MODTRAN5 radiative transfer code. When weather conditions were unsatisfactory, the total uncertainties of the original and improved methods were c. 5.9%–6.7% and c. 2.3%–3.5%, respectively, and the uncertainties caused by aerosol-type assumption were c. 15.8%–18.7% and c. 3.5%–8.0%, respectively. The results of the simulation and field experiments verified that the improved method has higher accuracy and lower uncertainty and is more suitable for the vicarious calibration of UAV hyperspectral remote sensors.

Description: Anomaly detection using sliding windows is not new but using causal sliding windows has not been explored in the past. The need of causality arises from real-time processing where the used sliding windows should not include future data samples that have not been visited, i.e., data samples come in after the currently being processed data sample. This paper develops an approach to anomaly detection using causal sliding windows, which has the capability of being implemented in real time. In doing so, three types of causal windows are defined: 1) causal sliding square matrix windows; 2) causal sliding rectangular matrix windows; and 3) causal sliding array windows. By virtue of causal sliding windows, a causal sample covariance/correlation matrix can be derived for causal anomaly detection. As for the causal sliding array windows, recursive update equations are also derived and thus speed up real-time processing.

Description: Segment-based classification is one of the popular approaches for object detection, where the performance of the classification task is sensitive to the accuracy of the output of the initial segmentation. Majority of the object detection systems directly use one of the generic segmentation algorithms, such as mean shift or k-means. However, depending on the problem domain, the properties of the regions such as size, color, texture, and shape, which are suitable for classification, may vary. Besides, fine tuning the segmentation parameters for a set of regions may not provide a globally acceptable solution in remote sensing domain, since the characteristic properties of a class in different regions may change due to the cultural and environmental factors. In this study, we propose a domain-specific segmentation method for building detection, which integrates information related to the building detection problem into the detection system during the segmentation step. Buildings in a remotely sensed image are distinguished from the highly cluttered background, mostly, by their rectangular shapes, roofing material and associated shadows. The proposed method fuses the information extracted from a set of unsupervised segmentation outputs together with this a priori information about the building object, called domain-specific information (DSI), during the segmentation process. Finally, the segmentation output is provided to a two-layer decision fusion algorithm for building detection. The advantage of domain-specific segmentation over the state-of-the-art methods is observed both quantitatively by measuring the segmentation and detection performances and qualitatively by visual inspection.

Description: Merging multiple satellite ocean color data is one of the ways to create a unified ocean color product and improve the spatial coverage. In this paper, the Bayesian maximum entropy (BME), a probabilistic method, is used to integrate chlorophyll-a (chl-a) concentration data obtained by the seaviewing wide field-of-view sensor (SeaWiFS) on Orbview-2, the medium-resolution imaging spectrometer instrument (MERIS) on ENVISAT and the moderate-resolution imaging spectroradiometer (MODIS) on Aqua. MODIS chl-a concentration on current day is considered as the accurate hard data. A probabilistic model is developed to link hard data and chl-a concentration of other sensors on previous days. The latter are processed as soft data by this probabilistic model to take into account the differences between mission-specific products. The semivariogram of chl-a concentration, which presents the spatial variability and provides a priori knowledge, is developed to improve the spatial coverage. The average daily coverage of the merged chl-a field is 74% for the 1-day temporal integration which is about six times higher than any single mission, and 95% for the 3-day temporal integration which achieves basically a complete global coverage. Root-mean-square error (RMSE) and correlation between in situ chl-a measurements and the BME-merged chl-a from 1-day data are 0.42 and 0.72, and from 3-day data are 0.44 and 0.70, respectively. Compared with the existing GSM method and the weighted averaging (AVW) method, the BME method can greatly improve the spatial coverage and preserve the high accuracy, which demonstrates the potential advantages of the BME method to merge ocean color products from multiple sensors.

Description: A novel method for calculating optimum incidence angle for the TanDEM-X system using any available digital elevation model (DEM) for the given area is proposed in this study. This method includes the plotting of slopes and aspect of the test area in a statistical way and applying mathematical approach using acquisition geometry in ascending and descending pass TanDEM-X data to optimize the incidence angle for obtaining precise DEM. Furthermore, the TanDEM-X raw DEMs in ascending and descending pass over Mumbai, India are combined using a simple weighted fusion algorithm and the quality of fused DEM thus generated is enhanced. The method adopted for fusion is just an experimental study. The problem of optimum weight selection for fusion has been addressed using height error map and a robust layover shadow mask calculated in “Integrated TanDEM-X Processor” (ITP) during TanDEM-X DEM generation. The height error map is calculated from the interferometric coherence with geometrical considerations and the robust layover and shadow map is calculated using TanDEM-X DEM and the corresponding slant range. Results show a significant reduction in the number of invalid pixels after fusion. In the fused DEM, invalids are only 2.14%, while ascending and descending pass DEMs have 5.02% and 6.34%, respectively. Statistical analysis shows a slight improvement in standard deviation of the error in fused DEM by 8% in urban area and about 5% for the whole scene. Only slight improvement in accuracy of fused DEM can be attributed to the coarse resolution of the SRTM-X DEM used as reference.

Description: The problem of determining and understanding the nature of buried objects by means of nondestructive and noninvasive techniques represents an interesting issue for a great variety of applications. In this framework, the theory of electromagnetic inverse scattering problems can help in such an issue by starting from the measures of the scattered field collected on a surface. What will be presented in this communication is a two-dimensional (2-D) technique based on the so-called Born approximation (BA) combined with a compressive sensing (CS) approach, in order to improve reconstruction capabilities for a proper class of targets. The use of a multiview-multistatic configuration will be employed together with a multifrequency approach to overcome the limited amount of data due to the single-frequency technique. Therefore, after a first numerical analysis of the performance of the considered algorithm, some numerical examples for 2-D aspect-limited configurations will be presented. The scenario is composed of a simplified scene, which consists of two half-spaces, and with the probes located close to the interface between the two media. As proposed in the following, it is easy to observe that the use of CS for this kind of problems may improve reconstruction capabilities, confirming the validity of the presented approach.

Description: Despite the active research, terrestrial laser scanning (TLS) has remained underutilized for forest structure assessment due to reliance of processing algorithms on high-resolution data, which may be costly and time-consuming to collect. Operational inventories, however, necessitate maximizing sample size while minimizing time and cost. The objective of this study was to assess the performance of a novel technique that enables stem reconstruction from low-resolution, single-scan TLS data in an effort to satisfy performance criteria against operational acquisition constraints. Instead of utilizing the curvature of the tree stem, e.g., by circle or cylinder fitting, we take advantage of the sensor-object geometry and reduce the dimensionality of the modeling to a series of one-dimensional (1-D) line fits. This allowed robust recovery of tree stem structure in a range of New England forest types, for tree stems which subtended at least an angular width of 15 mrad—the beam divergence of our system. Assessment was performed by projecting the three-dimensional (3-D) data onto two-dimensional (2-D) images and evaluating the per-point classification accuracies using manually digitized truth maps. Manual forest inventory measurements were also collected for each ${bf{20}} times {bf{20}};{bf{m}}$ plot and compared to measurements derived automatically. Good retrievals of stem location ( $R^2 = 0.99$ , ${bf{RMSE}} = {0.16;{bf{m}}}$ ) and diameter at breast height (DBH) ( $R^2 = {0.80}$ , ${bf{RMSE}} = {6.0;{bf{cm}}}$ ) were achieved. This study demonstrates that low-resolution sensors may be effective in pr- viding data for operational forest inventories constrained by sample size, time, and cost.

Description: Differential light detection and ranging (LiDAR) from repeated surveys has recently emerged as an effective tool to measure the three-dimensional (3-D) change. Currently, the primary method for determining 3-D change from LiDAR is through the use of the iterative closest point (ICP) algorithm and its variants, with a simplistic assumption of a uniform accuracy for the entire LiDAR point cloud. This common practice ignores the localization anisotropy and results in local convergence and spurious error estimation. To rigorously determine spatial change, this paper introduces an anisotropic-weighted ICP (A-ICP) algorithm, and proposes to model the random error for every LiDAR observation in the differential point cloud, and use this as a priori weights in the ICP algorithm. The implementation is evaluated by qualitatively and quantitatively comparing the estimation performance on point clouds with synthetic fault ruptures between standard ICP and A-ICP algorithm. As a further enhancement, we also present a moving window technique to improve A-ICP. Practical application of the combined moving window A-ICP technique is evaluated by estimating post-earthquake slip for the 2010 El Mayor-Cucapah Earthquake (EMC) using pre- and post-event LiDAR. Based on the analysis, moving window A-ICP is able to better estimate the synthetic surface ruptures, and provides a smoother estimate of actual displacement for the EMC earthquake.

Description: Earth observation hyperspectral imaging instruments capture and collect hundreds of different wavelength data corresponding to the same surface. As a result, tons of information must be stored, processed, and transmitted to ground by means of a combination of time-consuming processes. However, one of the requirements of paramount importance when dealing with applications that demand swift responses is the ability to achieve real-time. In this sense, the authors present a flexible and adaptable Field-Programmable Gate Array (FPGA)-based solution for extracting the endmembers of a hyperspectral image according to the Modified Vertex Component Analysis (MVCA) algorithm. The proposed approach is capable of adapting its parallelization execution by scaling the execution in hardware. Thus, the solution uses the dynamic and partial reconfiguration property of FPGAs in order to exploit and vary the level of parallelism at run-time. In order to validate the convenience of using this kind of solutions, the performance of our proposal has been assessed with a set of synthetic images as well as with the well-known Cuprite hyperspectral image. The achieved results demonstrate that the proposed system might be dynamically scaled without significantly affecting total execution times, being able to extract the endmembers of the Cuprite dataset in real-time.

Description: The high dimensionality of hyperspectral data constitutes a challenge for species classification. This study assessed 1) whether tree species classification can be optimized with the selection of bands which relate to known plant properties and 2) whether a partial least square (PLS) transformation improve species classification above principal component analysis (PCA). Leaf spectra between 400 and 2500 nm were measured for six evergreen tree species in the spring of 2011, in the KwaZulu-Natal Province of South Africa. Twenty-two bands which relate to pigment, foliage biomass, nutrients, and leaf structural components were selected from the hyperspectral data set. The 2100 bands of 1 nm were resampled to 421 bands at 5 nm spectral resolution, ensuring the number of variables are less than the number of samples. The random forest (RF) classification algorithm was used to assess the accuracy for both PCA and PLS transformations on the 421 and 22 bands. The accuracy of individual species classes was calculated as the average of ten iterations, for each data reduction option. The three 22-band models resulted in comparable accuracies to the 421-band classifications (OA of $mathbf{84} pm mathbf{4.9}% $ for untransformed, $mathbf{78} pm mathbf{5}% $ for PCA, and $mathbf{84} pm mathbf{4}% $ for PLS) and no significant differences between the 421 and 22-band models ( ${bm{p}} > mathbf{0.4}$ ). The optimized PLS model (22 bands, 8 components) showed a 6% ( ${bm{p}} 〈 mathbf{0.01}$ ) increase in accuracy above the optimized PCA model (22 bands, 3 components). Redu- ing hyperspectral data to bands which relate to plant properties, and the use of PLS for data transformation, optimizes species classification.

Description: This study assessed the suitability of chlorophyll, nitrogen, and water content, derived from leaf-level spectroradiometer data, for estimating volume of Eucalyptus clones in KwaZulu-Natal, South Africa. Volume was derived from field measurements of diameter at breast height ( dbh ) and tree height. Chlorophyll, nitrogen, and water related indices were used to estimate merchantable volume of Eucalyptus clones. Analysis of variance (ANOVA) was used to assess whether significant differences could be detected amongst index values within the plots or compartments, based on different age groups, clones, and site qualities. Cross validation and model selection based on adjusted ${mathrm{R}^2}$ and low Mallows’ Cp were utilized in the development of volume models. The strength of the correlations for all clones combined was found to be much lower than the individual relationships for E. grandis and E. saligna . ANOVA results indicated that volume was significantly ( $mathrm{p 〈 0.05}$ ) influenced by age, site quality, and the clone in question. Models developed using stepwise approach without ancillary data, such as age and site index, had low adjusted ${mathrm{R}^2}$ values ( $0.47 leq mathrm{R^2} leq 0.72$ ) and high root-mean-square error (RMSE) values compared to models that included ancillary data ( $0.81 leq mathrm{R^2} leq 0.90$ ). Partial least square regression models exhibited higher ${mathrm{R}^2}$ ( $0.92 leq mathrm{R^2} leq 0.96$ ) and lower RMSE and Mallows’ Cp. These results suggest that spectral measurements of chlorophyll, nitrogen, and water content have potential as independent variables to assist in the estimation of merchantable volume of Eucalyptus clones. This has important implications since results can be extended to airborne data and regional assessments.

Description: Precision agriculture requires detailed crop status information at high spatial and temporal resolutions. Remote sensing can provide such information, but single sensor observations are often incapable of meeting all data requirements. Spectral–temporal response surfaces (STRSs) provide continuous reflectance spectra at high temporal intervals. This is the first study to combine multispectral satellite imagery (from Formosat-2) with hyperspectral imagery acquired with an unmanned aerial vehicle (UAV) to construct STRS. This study presents a novel STRS methodology which uses Bayesian theory to impute missing spectral information in the multispectral imagery and introduces observation uncertainties into the interpolations. This new method is compared to two earlier published methods for constructing STRS: a direct interpolation of the original data and a direct interpolation along the temporal dimension after imputation along the spectral dimension. The STRS derived through all three methods are compared to field measured reflectance spectra, leaf area index (LAI), and canopy chlorophyll of potato plants. The results indicate that the proposed Bayesian approach has the highest correlation (r = 0.953) and lowest RMSE (0.032) to field spectral reflectance measurements. Although the optimized soil-adjusted vegetation index (OSAVI) obtained from all methods have similar correlations to field data, the modified chlorophyll absorption in reflectance index (MCARI) obtained from the Bayesian STRS outperform the other two methods. A correlation of 0.83 with LAI and 0.77 with canopy chlorophyll measurements are obtained, compared to correlations of 0.27 and 0.09, respectively, for the directly interpolated STRS.

Description: Sparse representation-based classification model has been widely applied into hyperspectral image (HSI) classification. Its mechanism is based on the assumption that the nonzero coefficients in the sparse representation mainly lie in the correct class-dependent low-dimensional subspace. However, the high similarity of pixels between some different classes exists in the HSI, which makes the classification process very unstable. In this paper, we propose a sparse representation based on the set-to-set distance (SRSTSD) for HSI classification. Through utilizing the set-to-set distance, the spatial information is incorporated into the sparse representation-based model. Moreover, to further exploit the spatial structure of the pixel, we also propose a patch-based SRSTSD (PSRSTSD) model. Experimental results demonstrate that our proposed methods can achieve excellent classification performance.

Description: Hyperspectral data classification is a hot topic in remote sensing community. In recent years, significant effort has been focused on this issue. However, most of the methods extract the features of original data in a shallow manner. In this paper, we introduce a deep learning approach into hyperspectral image classification. A new feature extraction (FE) and image classification framework are proposed for hyperspectral data analysis based on deep belief network (DBN). First, we verify the eligibility of restricted Boltzmann machine (RBM) and DBN by the following spectral information-based classification. Then, we propose a novel deep architecture, which combines the spectral–spatial FE and classification together to get high classification accuracy. The framework is a hybrid of principal component analysis (PCA), hierarchical learning-based FE, and logistic regression (LR). Experimental results with hyperspectral data indicate that the classifier provide competitive solution with the state-of-the-art methods. In addition, this paper reveals that deep learning system has huge potential for hyperspectral data classification.

Description: Sparse representation exhibits good performance in various image processing and has been applied to hyperspectral image (HSI) classification by many researchers. Recently, several new spatial–spectral strategies combined with sparse representation have been proposed to improve classification performance. However, these new strategies rely on spectral reflectance information and its neighborhood, without considering other spectral properties and higher order context information. Thus, in this paper, we present a spatial–spectral derivative-aided kernel joint sparse representation (KJSR-SSDK) for HSI classification. The proposed algorithm includes three novelties: 1) it considers the derivative features of the spectral as well as the original spectral feature; 2) it incorporates higher order spatial context and distinct spectral information; and 3) the $l_{1,2}$ mix-norm regularization is imposed on the coefficients of spatial–spectral derivative-aided dictionary for KJSR. Based on the rich experimental comparison with the related state-of-the-art algorithms, the effectiveness of the proposed KJSR-SSDK has been confirmed.

Description: The sparsity model has been employed for hyperspectral target detection and has been proved to be very effective when compared to the traditional linear mixture model. However, the state-of-art sparsity models usually represent a test sample via a sparse linear combination of both target and background training samples, which does not result in an efficient representation of a background test sample. In this paper, a sparse representation-based binary hypothesis (SRBBH) model employs more appropriate dictionaries with the binary hypothesis model to sparsely represent the test sample. Furthermore, the nonlinear issue is addressed in this paper, and a kernel method is employed to resolve the detection issue in complicated hyperspectral images. In this way, the kernel SRBBH model not only takes the nonlinear endmember mixture into consideration, but also fully exploits the sparsity model by the use of more reasonable dictionaries. The recovery process leads to a competition between the background and target subspaces, which is effective in separating the targets from the background, thereby enhancing the detection performance.

Description: In this paper, we propose a hyperspectral image anomaly detection model by the use of background joint sparse representation (BJSR). With a practical binary hypothesis test model, the proposed approach consists of the following steps. The adaptive orthogonal background complementary subspace is first estimated by the BJSR, which adaptively selects the most representative background bases for the local region. An unsupervised adaptive subspace detection method is then proposed to suppress the background and simultaneously highlight the anomaly component. The experimental results confirm that the proposed algorithm obtains a desirable detection performance and outperforms the classical RX-based anomaly detectors and the orthogonal subspace projection-based detectors.

Description: In this work, we assess the detection and classification of specially constructed targets in coincident airborne hyperspectral imagery (HSI) and high spatial resolution panchromatic imagery (HRI) in spectral, spatial, and joint spatial–spectral feature spaces. The target discrimination powers of the data-level and feature-level fusion of HSI and HRI are also directly compared in the spatial–spectral context using airborne imagery collected explicitly for this research. We show that in the case of Bobcat 2013 imagery, feature-level fusion of the HSI spectrum with spatial features derived from the coincident HRI data consistently results in fewer false alarms on the scene background as well as fewer misclassifications among the tested targets. Furthermore, this approach also outperforms schemes in which data-level fusion of the HSI and HRI imagery is performed prior to extracting spatial–spectral features.

Description: Constrained energy minimization (CEM) has been widely used in subpixel detection. This paper presents a new real-time processing of CEM according to two data acquisition formats, band-interleaved-pixel/sample (BIP/BIS) and band-interleaved line (BIL) where the global sample correlation matrix R must be replaced with a causal sample correlation matrix formed by only those data samples up to the pixel/sample currently being processed or a causal data line matrix formed by all data lines up to the data line being just completed. Both versions of CEM have not been investigated in the past. Its applications include detection of moving targets which can be only detected in real-time ongoing process, as well as subtle targets which is likely to be missed and overwhelmed by CEM in one-shot operation. Interestingly, while BIP/BIS and BIL look similar, their real-time implementations are quite different due to their use of causal sample correlation matrices, one updated by samples and the other updated by data lines. As a result, two different recursive equations are also derived for CEM using BIP/BIS and BIL, respectively, for real-time implementation.

Description: This paper investigates the retrieval of forest crown closure (CC) from the Landsat Thematic Mapper (TM) data and aerial images with a linear spectral mixture analysis (SMA) method. Anshan is selected as the study area. Two endmember extraction methods were used in this paper: 1) traditional image-based method and 2) up-scaling method. (When we get the fractions of components from a coregistered 0.6-m spatial resolution image, the linear spectral mixture model is applied to unmix the TM image and obtain the required endmembers.) For both methods, four fraction images (sunlit canopy, shaded canopy, sunlit background, shaded background) were calculated by linear spectral mixture model and used to derive CC. Results showed that CC can be fitted best with sum of fractions of sunlit canopy and shaded canopy at S-shaped curve and the up-scaling endmember extraction method is better than traditional image-based endmember extraction method. Finally, the up-scaling endmember extraction method was used to map forest CC in Anshan forested region. The measured forest CC distribution map was used to validate the estimated map. Results show that the estimated CC and measured CC have little difference and the estimated CC is slightly lower. The majority of Anshan forest CC values were between 0.4 and 0.8.

Description: Spectral unmixing is an important technique in hyperspectral image exploitation. It comprises the extraction of a set of pure spectral signatures (called endmembers in hyperspectral jargon) and their corresponding fractional abundances in each pixel of the scene. Over the last few years, many approaches have been proposed to automatically extract endmembers, which is a critical step of the spectral unmixing chain. Recently, ant colony optimization (ACO) techniques have reformulated the endmember extraction issue as a combinatorial optimization problem. Due to the huge computation load involved, how to provide suitable candidate endmembers for ACO is particularly important, but this aspect has never been discussed before in the literature. In this paper, we illustrate the capacity of ACO techniques for integrating the results obtained by different endmember extraction algorithms. Our experimental results, conducted using several state-of-the-art endmember extraction approaches using both simulated and a real hyperspectral scene (cuprite), indicate that the proposed ACO-based strategy can provide endmembers which are robust against noise and outliers.

Description: Spectral–spatial classification methods have been proven to be effective in hyperspectral image (HSI) classification. However, most of the methods make use of the correlation in a small neighborhood. In this paper, a novel low-rank decomposition spectral–spatial method (LRDSS) is proposed. LRDSS incorporates the global and local correlation where the global correlation is introduced by discovering the low-dimensional structure in the high-dimensional data, and local correlation is modeled by Markov Random Field (MRF). Specifically, all pixels’ spectrums in a homogeneous area are assumed to have low-dimensional structure. Low rankness is a fine property to characterize the low-dimensional structure and robust principal component analysis (RPCA) is used to extract the low-rank data. Then, the spectral information is obtained by the probabilistic support vector machine (SVM) classifier applied on the low-rank data. Moreover, the MRF models local correlation by encouraging neighboring pixels taking the same label. The maximum a posterior classification is computed by min-cut-based optimization algorithm. The experimental results suggest that LRDSS outperforms the other spectral–spatial classification methods investigated in this paper in terms of classification accuracies.

Description: The normal compositional model (NCM) has been introduced to characterize mixed pixels in hyperspectral images, particularly when endmember variability needs to be considered in the unmixing process. Each pixel is modeled as a linear combination of endmembers, which are treated as Gaussian random variables in order to capture such spectral variability. Since the combination coefficients (i.e., abundances) and the endmembers are unknown variables at the same time in the NCM, the parameter estimation is more difficult in comparison with conventional approaches. In order to address this issue, we propose a new Bayesian method, termed normal endmember spectral unmixing (NESU), for improved parameter estimation in this context. It considers the endmembers as known variables (resulting from the extraction of endmember bundles), then performs optimal estimations of the remaining unknown parameters, i.e., the abundances, using Bayesian inference. The particle swarm optimization (PSO) technique is adopted to estimate the optimal values of abundances according to their posterior probabilities. The performance of the proposed algorithm is evaluated using both synthetic and real hyperspectral data. The obtained results demonstrate that the proposed method leads to significant improvements in terms of unmixing accuracies.

Description: With the development of imaging technology, remote sensing images with a high spatial and spectral resolution have become available and have been used in various applications such as the identification of materials and the estimation of physical parameters. Although many endmember extraction algorithms have been proposed for hyperspectral data sets which extract/select the standard endmember spectrum for each existing endmember class or scene component, there are still some problems in endmember extraction which blur the discrimination between the different types of ground objects and lead to inaccurate endmember extraction. One problem is that the definition of pure materials (or endmembers) can be subjective and application dependent. The other problem is that spectral variability is inevitable due to the different imaging conditions, especially in a hyperspectral image with a higher spatial resolution. In this paper, to account for the spectral variability, each endmember of a material is represented with a set or “bundle” of spectra, and an image-based endmember bundle extraction algorithm using both spatial and spectral information is proposed. There are four steps in the proposed method of extracting endmember bundles: 1) pixel purity index preprocessing; 2) homogeneity index calculation; 3) region-based candidate endmember selection; and 4) spectral clustering. Experiments with both synthetic and real hyperspectral data sets indicate that, by considering the endmember variability in the original hyperspectral data, the proposed method shows a significant improvement over the current state-of-the-art endmember bundle extraction methods.

Description: Hyperspectral unmixing is an important step for material classification and recognition. Recently, nonnegative matrix factorization (NMF) has been utilized to unmix the hyperspectal imagery due to the advantage that it needs no assumption for the presence of pure pixels and can determine the endmembers and abundances simultaneously. In order to improve its unmixing performance, sparsity-constrained NMF has been demonstrated to be an efficient approach. The very recent study on $L_{1/2}$ regularization theory in compressive sensing (CS) and sparsity-constrained NMF show that the $L_{1/2}$ regularizer can yield stronger sparsity-promoting solutions than $L_{1}$ or $L_{2}$ regularizer. However, the $L_{1/2}$ regularization can result in a complex nonconvex optimization problem that is hard to solve efficiently. In this paper, we propose a fast and efficient adaptive half-thresholding algorithm for hyperspectral unmixing based on $L_{1/2}$ sparsity-constrained NMF. In the proposed algorithm, iterative half-thresholding procedure that has been proved to be an efficient method for solving $L_{1/2}$ regularization problem in CS is hybridized with the multiplicative update rule of standard NMF to deal with the $L_{1/2}$ sparsity-constrained NMF, which can give sparser and better unmixing results than the alternative algorithms. Furthermore, the data sparsity information can be incorporated into the algorithm to adap- ively adjust the regularization parameter of the model to improve algorithm performance and usability. The effectiveness of proposed algorithm was demonstrated by comparing with the representative algorithms on synthetic and real hyperspectral data.

Description: As a widely concerned research topic, many advanced algorithms have been proposed for hyperspectral unmixing. However, they may fail to accurately identify endmember signatures when coming across insufficient spatial resolution. To deal with this problem, an algorithm based on semisupervised linear sparse regression is proposed, in which unmixing procedure is reduced to seeking an optimal subset from the spectral library to best model mixed pixels in the scene. However, the number of the spectra with nonzero abundance is much more than that of the true endmember signatures. Furthermore, the selection of library spectra as endmember signatures is undesirable due to the divergent imaging conditions. In this paper, a novel projection-based nonnegative matrix factorization (NMF) (PNMF) algorithm is proposed by importing spectra library into the NMF framework. The main novelties of this paper are listed as follows. 1) By introducing the spectral library, the extraction of endmember signatures is no longer restricted by spatial resolution. 2) Related spectra are selected and projected onto a subspace containing the endmember signatures. So that the number of endmember signatures is controlled by dimension of the subspace. 3) In PNMF, the endmember signatures are adaptively generated from the spectral library, and are matched with the observed hyperspectral images. This overcomes the difficulty caused by diverse imaging conditions, and makes the proposed algorithm more practical for real applications. The experimental results, conducted on both synthetic and real hyperspectral data, illustrate the advantages of the proposed algorithm when compared with the state-of-the-art algorithms.

Description: This paper aims at defining a new paradigm (hypersharpening) in remote sensing image fusion. In fact, due to the development of new instruments, thinking only in terms of pansharpening is reductive. Even though some expressions as hyperspectral (HS) pansharpening already exist, there is not a suitable definition when multispectral/hyperspectral data are used as source to extract spatial details. After defining the hypersharpening framework, we draw the readers’ attention to its peculiar characteristics, by proposing and evaluating two hypersharpening methods. Experiments are carried out on the data produced by the updated version of SIM-GA imager, designed by Selex ES, which is composed by a panchromatic camera and two spectrometers in the VNIR and SWIR spectral ranges, respectively. Owing to the different resolution factors among panchromatic, VNIR and SWIR data sets, we can apply hypersharpening to fuse SWIR data to VNIR resolution. Comparisons of hypersharpening with “traditional” pansharpening show hypersharpening is more effective.

Description: Most satellites for environmental monitoring point ground targets from a perpendicular position to the terrain, also known as nadir. The CHRIS/PROBA platform provides four extra angles of observation, plus nadir pointing: $- mathbf{55}^{circ}$ , $- mathbf{36}^{circ}$ , $mathbf{0}^{circ}$ , $+ mathbf{36}^{circ}$ , and $+ mathbf{55}^{circ}$ . The purpose of this study is to investigate the best season (wet or dry season) and angle of observation to map tropical dry forest succession in Brazil. Nonparametric decision trees were used to build up classification maps based on principal component analysis (PCA) inputs. Our results indicate that the use of off-nadir data improves the map accuracy of successional stages of tropical dry forests and riparian forests. Particularly, extreme and negative angles of observation generate higher map accuracies, suggesting that tree shadows are enhancing spectral differences between the studied vegetation classes. Images from the dry season provide better total and classes’ map accuracies for late and intermediate stages of tropical dry forests. On the other hand, some classes, such as riparian forests and early stage of tropical forests need the use of off-nadir angles of observation to reach a minimum accuracy and best scores are reached using wet season’s images.

Description: Leaf nitrogen content has so far been quantified through empirical techniques using hyperspectral remote sensing. However, it remains a challenge to estimate the nitrogen content in fresh leaves through inversion of physically based models. Leaf nitrogen has been found to correlate with leaf traits (e.g., leaf chlorophyll, dry matter, and water) well through links to the photosynthetic process, which provides potential to estimate nitrogen indirectly. We therefore set out to estimate leaf nitrogen content by using its links to leaf traits that could be retrieved from a physically based model (PROSPECT) inversion. Leaf optical (directional-hemispherical reflectance and transmittance between 350 and 2500 nm) and leaf biochemical (nitrogen, chlorophyll, dry matter, and water) properties were measured. Correlation analysis showed that the area-based nitrogen correlations with leaf traits were higher than mass-based correlations. Hence, simple and multiple linear regression models were established for area-based nitrogen using three leaf traits (leaf chlorophyll content, leaf mass per area, and equivalent water thickness). In addition, the traits were retrieved by the inversion of PROSPECT using an iterative optimization algorithm. The established empirical models and the leaf traits retrieved from PROSPECT were used to estimate leaf nitrogen content. A simple linear regression model using only retrieved equivalent water thickness as a predictor produced the most accurate estimation of nitrogen ( ${R^2} = 0.58$ , $text{normalized};text{RMSE} = 0.11$ ). The combination of empirical and physically based models provides a moderately accurate estimation of leaf nitrogen content, which can be transferred to other datasets in a robust and upscalable manner.

Description: The efficient inversion of complex, three-dimensional (3-D) radiative transfer models (RTMs), such as the discrete anisotropy radiative transfer (DART) model, can be achieved using a look-up table (LUT) approach. A pressing research priority in LUT-based inversion for a 3-D model is to determine the optimal LUT grid size and density. We present a simple and computationally efficient approach for populating an LUT database with DART simulations over a large number of spectral bands. In the first step, we built a preliminary LUT using model parameters with coarse increments to simulate reflectance for six broad bands of Landsat Thematic Mapper (TM). In the second step, the preliminary LUT was compared with the TM reflectance, and the optimal input ranges and realistic parameter combinations that led to simulations close to Landsat spectra were then identified. In the third step, this information was combined with a sensitivity study, and final LUTs were built for the full spectrum of Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) narrow bands and six Landsat broad bands. The final LUT was inverted to estimate leaf area index (LAI) in northern temperate forests from AVIRIS and TM data. The results indicate that the approach used in this study can be a useful strategy to estimate LAI accurately by DART model inversion.

Description: Distance functions are at the core of important data analysis and processing tools, e.g., PCA, classification, vector median filter, and mathematical morphology. Despite its key role, a distance function is often used without careful consideration of its underlying assumptions and mathematical construction. With the objective of identifying a suitable distance function for hyperspectral images so as to maintain the accuracy of hyperspectral image processing results, we compare existing distance functions and define a suitable set of selection criteria. Bearing in mind that the selection of distance functions is highly related to the actual definition of the spectrum, we also classify the existing distance functions based on how they inherently define a spectrum. Theoretical constraints and behavior, as well as numerical tests are proposed for the evaluation of distance functions. With regards to the evaluation criteria, Euclidean distance of cumulative spectrum (ECS) was found to be the most suitable distance function.

Description: Aboveground biomass (AGB) of terrestrial ecosystems is an important constraint of global change and productivity models and used to assess carbon stocks and thus the contribution of vegetated ecosystems to the global carbon cycle. Although an indispensable and important requirement for decision makers, coherent and accurate estimates of grassland and forest AGB especially in complex environments are still lacking. In this study, we aim to assess the capability of two strategies to map grassland and forest AGB in a complex alpine ecosystem, i.e., using a discrete as well as a continuous field (CF) mapping approach based on imaging spectroscopy (IS) data. In situ measurements of grassland and forest AGB were acquired in the Swiss National Park (SNP) to calibrate empirical models and to validate AGB retrievals. The selection of robust empirical models considered all potential two narrow-band combinations of the simple ratio (SR) and the normalized difference vegetation index (NDVI) generated from Airborne Prism Experiment (APEX) IS data and in situ measurements. We found a narrow-band SR including spectral bands from the short-wave infrared (SWIR) (1689 nm) and near infrared (NIR) (851 nm) as the best regression model to estimate grassland AGB. Forest AGB showed highest correlation with an SR generated from two spectral bands in the SWIR (1498, 2112 nm). The applied accuracy assessment revealed good results for estimated grassland AGB using the discrete mapping approach [ ${mathrm{R}^2}$ of 0.65, mean RMSE (mRMSE) of $0.91,mathrm{t} cdot mathrm{h}{mathrm{a}^{ - 1}}$ , and mean relative RMSE (mrRMSE) of 26%]. The CF mapping approach produced a higher ${mathrm{R}^2}$ - /inline-formula〉 ( ${mathrm{R}^2} = 0.94$ ), and decreased the mRMSE and the mrRMSE to $0.55,mathrm{t} cdot mathrm{h}{mathrm{a}^{ - 1}}$ and 15%, respectively. For forest, the discrete approach predicted AGB with an ${mathrm{R}^2}$ value of 0.64, an mRMSE of $67.8,mathrm{t} cdot mathrm{h}{mathrm{a}^{ - 1}}$ , and an mrRMSE of 25%. The CF mapping approach improved the accuracy of forest AGB estimation with ${mathrm{R}^2} = 0.85$ , mean $mathrm{RMSE} = 55.85,mathrm{t} cdot mathrm{h}{mathrm{a}^{ - 1}}$ , and mean relative RMSE = 21%. Our results indicate that, in general, both mapping approaches are capable of accurately mapping grassland and forest AGB in complex environments using IS data, whereas the CF-based approach yielded higher accuracies due to its capability to incorporate subpixel information (abundances) of different land cover types.

Description: This paper presents a novel and efficient spectral–spatial classification method for hyperspectral images. It combines the spectral and texture features to improve the classification accuracy. The moment invariants are computed within a small window centered at the pixel to determine pixel-wise texture features. The texture and spectral features are concatenated to form a joint feature vector that is used for classification with support vector machine (SVM). The experiments are carried out on three hyperspectral datasets and results are compared with some other spectral–spatial techniques. The results indicate that the proposed method statistically significantly improved the classification accuracies over the conventional spectral method. The new method has also outperformed the other recently used spectral–spatial methods in terms of both classification accuracies and computational cost. The results also showed that the proposed method can produce good classification accuracy with smaller training sets.

Description: In land cover assessment, classes often gradually change from one to another. Therefore, it is difficult to allocate sharp boundaries between different classes of interest. To overcome this issue and model such conditions, fuzzy techniques that resemble human reasoning have been proposed as alternatives. Fuzzy C-means is the most common fuzzy clustering technique, but its concept is based on a local search mechanism and its convergence rate is rather slow, especially considering high-dimensional problems (e.g., in processing of hyperspectral images). Here, in order to address those shortcomings of hard approaches, a new approach is proposed, i.e., fuzzy C-means which is optimized by fractional order Darwinian particle swarm optimization. In addition, to speed up the clustering process, the histogram of image intensities is used during the clustering process instead of the raw image data. Furthermore, the proposed clustering approach is combined with support vector machine classification to accurately classify hyperspectral images. The new classification framework is applied on two well-known hyperspectral data sets; Indian Pines and Salinas. Experimental results confirm that the proposed swarm-based clustering approach can group hyperspectral images accurately in a time-efficient manner compared to other existing clustering techniques.

Description: Spectral variation, which is inevitably present in hyperspectral data due to nonuniformity and inconsistency of illumination, may result in considerable difficulty in spectral unmixing. In this paper, a field endmember library is constructed to accommodate spectral variation by representing each endmember class by a batch of image-derived spectra. In order to perform unmixing by such a field endmember library, a novel spatial and spectral endmember selection (SSES) algorithm is designed to search for a spatial and spectral constrained endmember subset per pixel for abundance estimation (AE). The net effect is to achieve sparse unmixing equivalently, considering the fact that only a few endmembers in the large library have nonzero abundances. Thus, the resulting algorithm is called spatial and spectral constrained sparse unmixing (SSCSU). Experimental results using both synthetic and real hyperspectral images demonstrate that the proposed SSCSU algorithm not only improves the performance of traditional AE algorithms by considering spectral variation, but also outperforms the existing sparse unmixing approaches.

Description: Nonlinear hyperspectral unmixing (HSU) plays a key-role in understanding and quantifying the physical-chemical phenomena occurring over geometrically complex fields of view. Nonlinear HSU methods that do not rely on prior knowledge of the ground truth to analyze the scene are especially interesting. However, they can be affected either by overfitting or performance degradation provided by inaccurate setting of unmixing parameters. In this paper, we introduce a new nonlinear HSU architecture which aims at taking advantage of the benefit provided by the combination of polytope decomposition (POD) method together with artificial neural network (ANN)-based learning. Specifically, ANN is able to efficiently estimate the order $p$ of the nonlinearity provided by the given scene even without the thorough knowledge of the ground truth. The ANN-based learning is used to feed the POD in order to deliver accurate unmixing based on a $p$ -linear polynomial model. Experimental results over simulated and real scenes show promising performance of the proposed framework.

Description: This paper develops a new approach, called progressive band processing (PBP) of linear spectral unmixing (LSU) which can process data unmixing according to band sequential (BSQ) format. This new concept is quite different from band selection (BS) which must select bands from a fully collected band set based on a band optimization criterion. There are several advantages of using PBP-LSU over BS. In particular, it allows users to perform LSU using available bands without waiting for a complete collection of full bands. In doing so, an innovations information update equation is further derived and can process LSU as its band process is ongoing. To be more specific, LSU can be carried out by PBP by updating unmixed data recursively band by band in the same way that a Kalman filter does for updating data information in a recursive fashion. With such a recursion in nature, PBP is able to process bands of interest which may vary with different applications.

Description: Unmixing provides a summary of hyperspectral data and is useful for many image processing tasks. Recently, spectral unmixing has also been introduced to hyperspectral image denoising literature. However, so far, only the spectral information has been utilized for unmixing-based denoising. While most of the endmember extraction methods in the literature rely solely on spectral information, it has been shown that spatial-spectral preprocessing (SSPP) methods can enhance endmember extraction performance by utilizing the assumption that endmembers are more likely to be located in homogenous regions. This letter proposes the use of SSPP prior to spectral unmixing, to guide the endmember extraction process to spatially homogenous regions. The enhanced endmember extraction performance in turn leads to enhanced denoising performance. In addition, the proposed approach also goes one step further and retains the anomalous/scarce endmembers, which may include important endmembers, such as rare minerals, stressed crops, or military targets, and which may be lost due to the included spatial preprocessing (SPP) steps. Discarding such anomalous endmembers in a summary or compression of big data may result in undesired consequences. In short, the proposed approach provides enhanced unmixing-based denoising performance, while also retaining the anomalous endmembers.

Description: In the available hyperspectral unmixing approaches, hyperspectral images are often treated as a list of spectral measurements with no geometric organization. In this paper, we explore the geometric structure of hyperspectral images in both the spatial domain and the spectral domain to advance a geometric nonnegative matrix factorization (GNMF) for more accurate endmember extraction and abundance estimation. In the proposed GNMF, we define the “spatial geometric distance” and “spectral geometric distance” to reveal the affinity between pixels. Both the spatial geometric homogeneity of hyperspectral vectors in a local region, and the geometry manifold structure in the spectral domain, are explored to formulate spatial–spectral manifold regularizer for NMF. Some experiments are taken on some synthetic data and real hyperspectral data to investigate the performance of GNMF, and the results show that it can present state-of-the-art unmixing results.

Description: In this paper, a novel multiple kernel learning (MKL) algorithm is proposed for the classification of hyperspectral images. The proposed MKL algorithm adopts a two-step strategy to learn a multiple kernel machine. In the first step, unsupervised learning is carried out to learn a combined kernel from the predefined base kernels. In our algorithms, low-rank nonnegative matrix factorization (NMF) is used to carry out the unsupervised learning and learn an optimal combined kernel. Furthermore, the kernel NMF (KNMF) is introduced to substitute NMF for enhancing the ability of the unsupervised learning with the predefined base kernels. In the second step, the optimal kernel is embedded into the standard optimization routine of support vector machine (SVM). In addition, we address a major challenge in hyperspectral data classification, i.e., using very few labeled samples in a high-dimensional space. Experiments are conducted on three real hyperspectral datasets, and the experimental results show that the proposed algorithms, especially for KNMF-based MKL, achieve the outstanding performance for hyperspectral image classification with few labeled samples when compared with several state-of-the-art algorithms.

Description: Pixel purity index (PPI) and N-finder algorithm (N-FINDR) are the two most widely used techniques for finding endmembers in hyperspectral imagery. Unfortunately, both of them suffer from two same major issues, inconsistency resulting from random initial conditions and computational complexity resulting from finding all endmembers simultaneously. To resolve these issues, a theory of iterative N-FINDR (IN-FINDR) has been recently developed for N-FINDR via iterative processes. This paper develops various versions of iterative PPI (IPPI), which can be considered as a companion paper of IN-FINDR. The IPPI reinvents a wheel by redesigning various versions of iterative algorithms which implement PPI via two iterative processes, one for data sample vectors and the other for skewers. More specifically, two different iterative versions of IPPI are developed, called progressive IPPI (P-IPPI) and causal IPPI (C-IPPI), which correspond to two versions of IN-FINDR, successive N-FINDR (SC N-FINDR) and sequential N-FINDR (SQ N-FINDR), respectively. Because they are iterative in nature, C-IPPI and P-IPPI can be further extended to process skewer sets in two different fashions, varying skewer set and growing skewer set, to be called varying skewer set C-IPPI (VC-IPPI) and growing skewer set P-IPPI (GP-IPPI). As a result, the fast IPPI (FIPPI) can be shown indeed to be their special case.

Description: A new concept of virtual signature (VS) is introduced for linear spectral mixture analysis (LSMA) that can be used to form a linear mixture model (LMM) to perform data unmixing. A VS is defined as a spectrally distinct signature and different from an endmember as a pure signature in the sense that a VS must be extracted directly from the data and can be a mixed signature. It is also different from virtual endmembers (VEs) introduced in nonlinear spectral mixture analysis according to a bilinear model. By virtue of VS, this paper investigates three least squares (LSs)-based criteria, LS error (LSE), orthogonal subspace projection (OSP) residual, and maximal likelihood estimation (MLE) error, and further designs of their respective recursive algorithms to find VSs for LSMA to unmix data. In the mean time, a binary composite hypothesis testing-based Neyman Pearson detector is also developed in conjunction with the developed recursive algorithms to determine if their produced signatures are indeed desired VSs. Finally, an experiment-based comparative analysis is also conducted to demonstrate the proposed approaches.

Description: A nonlinear joint collaborative representation (CR) model with adaptive weighted multiple feature learning to deal with the small sample set problem in hyperspectral image (HSI) classification is proposed. The proposed algorithm first maps every meaningful feature of the image scene into a kernel space by a column-generation (CG)-based technique. A unified multitask learning-based joint CR framework, with adaptive weighting for each feature, is then undertaken by the use of an alternating optimization algorithm, to obtain accurate kernel representation coefficients, which leads to desirable classification results. The experimental results indicate that the proposed algorithm obtains a competitive performance and outperforms the other state-of-the-art regression-based classifiers and the classical support vector machine classifier.

Description: Imaging spectrometers can supply spatial data in abundant narrow and continuous wavelength bands. However, the huge data volume produced encounters the difficulty in storage and transmission. On the other hand, these hyperspectral data sets contain high redundancy, which offers an opportunity to reduce the number of spectral measurements and recover the full spectrum from limited samples without losing principal spectral information. This paper addresses the application of compressed sensing (CS) theory to hyperspectral data reconstruction. An important question involved is how to know a spectrum is sparse enough so that CS can be applied effectively. We provide a quantitative answer and develop a strategy to measure the degree of sparsity of a spectrum based on the generalized Gaussian distribution (GGD) model. The novelty includes the derivation of the sharpness of the GGD and how to estimate the sharpness of a spectral signal. The proposed strategy was tested using the spectral data from USGS database and an AVIRIS-HSI data set. The results demonstrate that it is important to introduce the sparsity measure, as CS offers a high reconstruction rate and low relative errors compared with the existing methods for sparse signals only.

Description: An improved sparse subspace clustering (ISSC) method is proposed to select an appropriate band subset for hyperspectral imagery (HSI) classification. The ISSC assumes that band vectors are sampled from a union of low-dimensional orthogonal subspaces and each band can be sparsely represented as a linear or affine combination of other bands within its subspace. First, the ISSC represents band vectors with sparse coefficient vectors by solving the L2-norm optimization problem using the least square regression (LSR) algorithm. The sparse and block diagonal structure of the coefficient matrix from LSR leads to correct segmentation of band vectors. Second, the angular similarity measurement is presented and utilized to construct the similarity matrix. Third, the distribution compactness (DC) plot algorithm is used to estimate an appropriate size of the band subset. Finally, spectral clustering is implemented to segment the similarity matrix and the desired ISSC band subset is found. Four groups of experiments on three widely used HSI datasets are performed to test the performance of ISSC for selecting bands in classification. In addition, the following six state-of-the-art band selection methods are used to make comparisons: linear constrained minimum variance-based band correlation constraint (LCMV-BCC), affinity propagation (AP), spectral information divergence (SID), maximum-variance principal component analysis (MVPCA), sparse representation-based band selection (SpaBS), and sparse nonnegative matrix factorization (SNMF). Experimental results show that the ISSC has the second shortest computational time and also outperforms the other six methods in classification accuracy when using an appropriate band number obtained by the DC plot algorithm.

Description: The high dimensionality of hyperspectral images brings a heavy burden for image processing. Band selection is a common technique for dimensionality reduction. Since the labels of hyperspectral images are difficult to collect, a new semisupervised band selection method based on affinity propagation (AP) is proposed. AP, an exemplar-based clustering method, is famous due to fast execution time and low reconstruction error. For band selection, AP involves two key issues: band correlation and band preference. In this paper, a new normalized trivariable mutual information (normalized TMI, NTMI) is devised to measure band correlation for classification. NTMI considers not only band redundancy but also band synergy, and overcomes the sensitivity of TMI to the discriminative abilities of bands. Band preference is defined by the discriminative ability and informative amount of each band. Since the clustering methods are easily disturbed by noisy bands, a new statistical-based method for band correlation and band preference is devised. It can automatically remove noisy bands beforehand by exploiting the continuity property of bands. Finally, the proposed method can select highly discriminative and informative bands, and remove highly redundant bands. Experimental results on hyperspectral images demonstrate the effectiveness of the proposed semisupervised band selection method.

Description: In spectral analysis, diagnostic absorption features can indicate the existence of specific materials. Absorption parameters such as absorption center, absorption width, and absorption depth can be used in not only identification and quantitative analysis of minerals, but also in retrieval of surface physical properties. Continuum removal (CR) is commonly used to extract absorption features. However, for a band range containing more than one absorption contribution factors, the feature extracted by CR could be a result of comprehensive effect of different factors. In this paper, a new spectral feature extraction method named reference spectral background removal (RSBR) is proposed. Given the reference spectral background, RSBR can eliminate the influence of unwanted contribution factor, and extract the absorption feature of target contribution factor. Using RSBR, the basic absorption feature parameters including the absorption center, absorption width, and absorption depth are extracted. The results are compared with those obtained from the CR. It is shown that RSBR can effectively extract pure absorption features of target material, while more accurate absorption parameters can also be achieved.

Description: As a recent approach for time series analysis, singular spectrum analysis (SSA) has been successfully applied for feature extraction in hyperspectral imaging (HSI), leading to increased accuracy in pixel-based classification tasks. However, one of the main drawbacks of conventional SSA in HSI is the extremely high computational complexity, where each pixel requires individual and complete singular value decomposition (SVD) analyses. To address this issue, a fast implementation of SSA (F-SSA) is proposed for efficient feature extraction in HSI. Rather than applying pixel-based SVD as conventional SSA does, the fast implementation only needs one SVD applied to a representative pixel, i.e., either the median or the mean spectral vector of the HSI hypercube. The result of SVD is employed as a unique transform matrix for all the pixels within the hypercube. As demonstrated in experiments using two well-known publicly available data sets, almost identical results are produced by the fast implementation in terms of accuracy of data classification, using the support vector machine (SVM) classifier. However, the overall computational complexity has been significantly reduced.

Description: This paper proposes a new approach of band selection for classifying multiple objects in hyperspectral images. Different from traditional algorithms, we construct a semisupervised pair-wise band selection (PWBS) framework for this task, in which an individual band selection process is performed only for each pair of classes. First, the statistical parameters for spectral features of each class, including mean vectors and covariance matrices, are estimated by an expectation maximization approach in a semisupervised learning setting, where both labeled and unlabeled samples are employed for better performance. For each pair of classes, based on the estimated statistical parameters, Bhattacharyya distances between the two classes are calculated to evaluate all possible subsets of bands for classification. Second, as our proposed semisupervised framework, the PWBS followed by a binary classifier can be embedded into the semisupervised expectation maximization process to obtain posterior probabilities of samples on the selected bands. Finally, to evaluate the selected bands, all of the binary decisions obtained with multiple binary classifiers are finally fused together. Comparative experimental results demonstrate the validity of our proposed algorithm. The experimental results also prove that our band selection algorithm can perform well when the training set is very small.