ALBERT

Description: Comparing the performance of different satellite sensors in global land cover change (LCC) monitoring is necessary to assess their potential and limitations for more accurate and operational LCC estimations. This paper aims to examine and compare the performance in LCC monitoring using three satellite sensors: PROBA-V, Landsat 8 OLI, and Sentinel-2 MSI. We utilized a unique set of global reference data containing four years of records (2015–2018) at 29,263 land cover change/no-change 100 × 100-m sites. The LCC monitoring was conducted using the BFAST(s)-Random Forest (BRF) change detection framework involving 15 global timeseries vegetation indices and three BFAST models. Due to the different spectral characteristics and data availability of the sensors, we designed 30 comparison scenarios to extensively evaluate their performance. The overall results were: 1) for global general LCC monitoring, Landsat 8 OLI slightly outperformed Sentinel-2, and PROBA-V performed the worst. The performance among the three sensors differed consistently despite different data availability and spectral observation regions. Sentinel-2 was more competitive with Landsat 8 when the red-edge 1 band was included; 2) Landsat 8 was more accurate in forest, herbaceous vegetation, and water monitoring. Sentinel-2 performed particularly well in wetland monitoring. In addition, we further observed: 3) missing data in time series decreased the accuracy in all sensors, but had little influence on the relative performance across sensors; 4) combining sensors would not necessarily improve the accuracy because the complementary effects enhanced the accuracy only when there was a large amount of data missing for all sensors; 5) the BRF framework maintained the performance gap among sensors, but obtained a higher and more balanced accuracy overall when compared with using BFAST methods alone, by involving ensemble learning with an embedded sample-balancing strategy; 6) among the random forest variables, the ‘magnitude’ proved to be the most important contributor, and the NDVI had the most consistently good performance across sensors when compared against other vegetation indices. All sensors using BRF still had some errors in change detection, with a tendency to underestimate the global LCC. A potential reason for this is the complexity of the diverse change/no-change characteristics at the global extent and the fact that smaller, more subtle LCCs might not be well detected. These limitations could be addressed by taking advantage of ensemble learning approaches with a combination of multiple independent region/thematic-adapted LCC monitoring models and using the original Sentinel-2 (10 m) and Landsat 8 (30 m) in the future.

Description: Currently most global land cover maps are produced with discrete classes, which express the dominant land cover class in each pixel, or a combination of several classes at a predetermined ratio. In contrast, land cover fraction mapping enables expressing the proportion of each pure class in each pixel, which increases precision and reduces legend complexity. To map land cover fractions, regression rather than classification algorithms are needed, and multiple approaches are available for this task. A major challenge for land cover fraction mapping models is data sparsity. Land cover fraction data is by its nature zero-inflated due to how common the 0% fraction is. As regression favours the mean, 0% and 100% fractions are difficult for regression models to predict accurately. We proposed a new solution by combining three models: a binary model determines whether a pixel is pure; if so, it is processed using a classification model; otherwise with a regression model. We compared multiple regression algorithms and implemented our proposed three-step model on the algorithm with the lowest RMSE. We further evaluated the spatial and per-class accuracy of the model and demonstrated a wall-to-wall prediction of seven land cover fractions over the globe. The models were trained on over 138,000 points and validated on a separate dataset of over 20,000 points, provided by the CGLS-LC100 project. Both datasets are global and aligned with the PROBA-V 100 m UTM grid. Results showed that the random forest regression model reached the lowest RMSE of 17.3%. Lowest MAE (7.9%) and highest overall accuracy (72% ± 2%) was achieved using random forest with our proposed three-model approach and median vote. This research proves that machine learning algorithms can be applied globally to map a wide variety of land cover fractions. Fraction mapping expresses land cover more precisely, and empowers users to create their own discrete maps using user-defined thresholds and rules, which enables customising the result for a diverse range of uses. The three-step approach is useful for addressing the zero-inflation issue and mapping 0% and 100% fractions more accurately, and thus has already been taken up in the operational production of global land cover fraction layers within the CGLS-LC100 project. Furthermore, this study contributes to the accuracy assessment of land cover fraction maps both thematically and spatially, and these methods could be taken up by future land cover fraction mapping efforts.

Description: BFAST Lite is a newly proposed unsupervised time series change detection algorithm that is derived from the original BFAST (Breaks for Additive Season and Trend) algorithm, focusing on improvements to speed and flexibility. The goal of the BFAST Lite algorithm is to aid the upscaling of BFAST for global land cover change detection. In this paper, we introduce and describe the algorithm and then compare its accuracy, speed and features with other algorithms in the BFAST family: BFAST and BFAST Monitor. We tested the three algorithms on an eleven-year-long time series of MODIS imagery, using a global reference dataset with over 30,000 point locations of land cover change to validate the results. We set the parameters of all algorithms to comparable values and analysed the algorithm accuracy over a range of time series ordered by the certainty of that the input time series has at least one abrupt break. To compare the algorithm accuracy, we analysed the time difference between the detected breaks and the reference data to obtain a confusion matrix and derived statistics from it. Lastly, we compared the processing speed of the algorithms using both the original R code as well as an optimised C++ implementation for each algorithm. The results showed that BFAST Lite has similar accuracy to BFAST but is significantly faster, more flexible and can handle missing values. Its ability to use alternative information criteria to select the number of breaks resulted in the best balance between the user’s and producer’s accuracy of detected changes of all the tested algorithms. Therefore, BFAST Lite is a useful addition to the BFAST family of unsupervised time series break detection algorithms, which can be used as an aid in narrowing down areas with changes for updating land cover maps, detecting disturbances or estimating magnitudes and rates of change over large areas.

Description: Annual global land cover maps (GLC) are being provided by several operational monitoring efforts. However, map validation is lagging, in the sense that the annual land cover maps are often not validated. Concurrently, users such as the climate and land management community require information on the temporal consistency of multi-date GLC maps and stability in their accuracy. In this study, we propose a framework for operational validation of annual global land cover maps using efficient means for updating validation datasets that allow timely map validation according to recommendations in the CEOS Stage-4 validation guidelines. The framework includes a regular update of a validation dataset and continuous map validation. For the regular update of a validation dataset, a partial revision of the validation dataset based on random and targeted rechecking (areas with a high probability of change) is proposed followed by additional validation data collection. For continuous map validation, an accuracy assessment of each map release is proposed including an assessment of stability in map accuracy addressing the user needs on the temporal consistency information of GLC map and map quality. This validation approach was applied to the validation of the Copernicus Global Land Service GLC product (CGLS-LC100). The CGLS-LC100 global validation dataset was updated from 2015 to 2019. The update was done through a partial revision of the validation dataset and an additional collection of sample validation sites. From the global validation dataset, a total of 40% (10% for each update year) was revisited, supplemented by a targeted revision focusing on validation sample locations that were identified as possibly changed using the BFAST time series algorithm. Additionally, 6720 sample sites were collected to represent possible land cover change areas within 2015 and 2019. Through this updating mechanism, we increased the sampling intensity of validation sample sites in possible land cover change areas within the period. Next, the dataset was used to validate the annual GLC maps of the CGLS-LC100 product for 2015–2019. The results showed that the CGLS-LC100 annual GLC maps have about 80% overall accuracy showing high temporal consistency in general. In terms of stability in class accuracy, herbaceous wetland class showed to be the least stable over the period. As more operational land cover monitoring efforts are upcoming, we emphasize the importance of updated map validation and recommend improving the current validation practices and guidelines towards operational map validation so that long-term land cover maps and their uncertainty information are well understood and properly used.