ALBERT

Description: Millions of tons of plastic waste are released into the marine environment every year. While they steadily accumulate, synthetic polymers provide a habitat for microorganisms. This denominated Plastisphere has been studied in detail over the past ten years. So have the enzymes responsible for microbial degradation, which are unfortunately lacking for most sorts of plastics. Therefore, the BMBF-funded project PLASTISEA is focusing on bioprospecting the marine treasure trove for novel plastic acting enzymes.

Description: Many of the socio-economic and environmental challenges of the 21st century like the growing energy and food demand, rising sea levels and temperatures put stress on marine ecosystems and coastal populations. This requires a significant strengthening of our monitoring capacities for processes in the water column, at the seafloor and in the subsurface. However, present-day seafloor instruments and the required infrastructure to operate these are expensive and inaccessible. We envision a future Internet of Underwater Things, composed of small and cheap but intelligent underwater nodes. Each node will be equipped with sensing, communication, and computing capabilities. Building on distributed event detection and cross-domain data fusion, such an Internet of Underwater Things will enable new applications. In this paper, we argue that to make this vision a reality, we need new methodologies for resource-efficient and distributed cross-domain data fusion. Resource-efficient, distributed neural networks will serve as data-analytics pipelines to derive highly aggregated patterns of interest from raw data. These will serve as (1) a common base in time and space for fusion of heterogeneous data, and (2) be sufficiently small to be transmitted efficiently in resource-constrained settings.

Description: Die Messung submariner Bodendeformationen an den Flanken von Inselvulkanen hilft dabei, ihre Stabilität und die Gefahr von Hangrutschungen einzuschätzen, ist aber inherent schwierig für Gebiete, die unter Wasser liegen. Wiederholte Seismik- oder Fächerecholot-Vermessungen können größere Gebiete abdecken, aber Auflösung und Lokalisierung sind bestimmten Grenzen unterworfen. Optische Daten andererseits sind besser aufgelöst, aber limitiert in ihrer räumlichen Abdeckung, und Meeresbodengeodäsie wiederum liefert nur punktuelle Information. In diesem Artikel schlagen wir vor, verschiedene Arten von Fernerkundungsdaten zusammenzubringen und auch mit bestehenden statischen und dynamischen Modellen zu verschneiden. Aufgrund ihrer verschiedenen Modalitäten, Unsicherheiten und Skalierungen ist dies jedoch schwierig und bedarf einer Fusion. Zusammen mit anderen Aspekten (Erdbeben, Strömungen etc.) sollen die fusionierten Daten und Modelle langfristig neue Einblicke in das dynamische System des sich verändernden Meeresbodens, die dafür verantwortlichen Faktoren sowie die Auswirkungen instabiler submariner Hänge auf andere Meeressysteme bieten.

Description: A central promise of cross-domain fusion (CDF) is the provision of a “bigger picture” that integrates different disciplines and may span very different levels of detail. We present a number of settings that call for this bigger picture, with a particular focus on how information from several domains can be made easily accessible and visualizable for different stakeholders. We propose harnessing an approach that is now well established in interactive maps, which we refer to as the “Google maps approach” (Google LLC, Mountain View, CA, USA), which combines effective filtering with intuitive user interaction. We expect this approach to be applicable to a range of CDF settings.

Description: Consistently high data quality is essential for the development of novel loss functions and architectures in the field of deep learning. The existence of such data and labels is usually presumed, while acquiring high-quality datasets is still a major issue in many cases. Subjective annotations by annotators often lead to ambiguous labels in real-world datasets. We propose a data-centric approach to relabel such ambiguous labels instead of implementing the handling of this issue in a neural network. A hard classification is by definition not enough to capture the real-world ambiguity of the data. Therefore, we propose our method “Data-Centric Classification & Clustering (DC3)” which combines semi-supervised classification and clustering. It automatically estimates the ambiguity of an image and performs a classification or clustering depending on that ambiguity. DC3 is general in nature so that it can be used in addition to many Semi-Supervised Learning (SSL) algorithms. On average, our approach yields a 7.6% better F1-Score for classifications and a 7.9% lower inner distance of clusters across multiple evaluated SSL algorithms and datasets. Most importantly, we give a proof-of-concept that the classifications and clusterings from DC3 are beneficial as proposals for the manual refinement of such ambiguous labels. Overall, a combination of SSL with our method DC3 can lead to better handling of ambiguous labels during the annotation process. (Source code is available at https://github.com/Emprime/dc3).

Description: Here, we outline how to identify hydrogenase enzymes from metagenomic fosmid libraries through an activity-based screening approach. A metagenomic fosmid library is constructed in E. coli and the fosmids are transferred into a hydrogenase deletion mutant of Shewanella oneidensis MR-1 (ΔhyaB) via triparental mating. If a fosmid clone exhibits hydrogen-uptake activity, S. oneidensis’ phenotype is restored and hydrogenase activity is indicated by a color change of the medium from yellow to colorless. The screen enables screening of 48 metagenomic fosmid clones in parallel.

Description: The gas hydrate system off Mauritania is characterized by the undulating landward termination of a gas hydrate-related bottom simulating reflector (BSR). Some of the most landward sections of this BSR reach up to within 6 m of the seafloor. This suggests a shallow sulphate-methane-interface over an unusually large area. We attribute this to the presence of large amounts of methane due to the efficient burial of organic matter in a high-productivity oceanographic region, and the efficient channelling of methane along permeable turbidite beds towards the feather edge of the gas hydrate stability zone. This is consistent with the observation of steps in the BSR, where it cross-cuts other inferred permeable horizons. The high thermal conductivity of a salt dome in the southern part of the study area distorts the subsurface temperature field, giving the base of the hydrate stability zone a concave-down shape. Within this anticline of the BSR, high amplitudes and a horizontal reflection that crosses the sedimentary strata indicate the entrapment of free gas. We interpret this as a direct indication of a reduced hydraulic permeability of the hydrate-bearing sediment.

Description: Spectacular advances have been made in the field of machine vision over the past decade. While this discipline is traditionally driven by geometric models, neural networks have proven to be superior in some applications and have significantly expanded the limits of what is possible. At the same time, conventional graphic models describe the relationship between images and the associated scene with textures and light in a physically realistic manner and are an important part of photogrammetry. Differential renderers combine these approaches by enabling gradient-based optimization in fixed structures of a graphics pipeline and thus adapt the learning process of neural networks. This fusion of formalized knowledge and machine learning motivates the idea of a modular differentiable renderer in which physical and statistical models can be recombined depending on the use case. We therefore present Gemini Connector: an initiative for the modular development and combination of differentiable physical models and neural networks. We examine opportunities and problems and motivate the idea with the extension of a differentiable rendering pipeline to include models of underwater optics for the analysis of deep sea images. Finally, we discuss use cases, especially within the Cross-Domain Fusion initiative.

Description: Interactive exploration of Earth system simulations may have great potential to improve the scientific modeling process. It will allow monitoring of the state of the simulation via dashboards presenting real-time diagnostics within a digital twin world. We present the state of the art for Earth system modeling in this context. Cross-domain data handling and fusion will make it possible to integrate model and observation data in the context of digital twins of the ocean. Domain-driven modularization of monolithic Earth system models allows one to recover interfaces for such a cross-domain fusion. Reverse engineering with static and dynamic analysis enables modularization of Earth system models. The modularization does not only help with restructuring existing Earth system models, it also makes it possible to integrate additional scientific domains into the interactive simulation environment.

Description: We explore the predictability of tropical Atlantic sea surface temperature (SST) and the potential influence of climate model bias on SST predictions over the tropical Atlantic. Two statistical methods are used to examine the skill in forecasting tropical Atlantic SST anomalies (SSTAs): linear inverse modeling (LIM) and analogue forecast (AF). The statistical models are trained either with observations or with data from two control integrations of the Kiel Climate Model (KCM), which only differ with respect to the resolution of its atmospheric component. Observed SSTAs suggest that Tropical Atlantic climatic changes are potentially predictable at lead times of up to 6 months over large parts of the Tropical Atlantic. The SSTAs from the KCM version employing a high-resolution atmosphere model (KCM-HRES) is potentially predictable at a level comparable to that derived from the observations, whereas the SSTAs from the KCM version employing a low-resolution atmosphere model (KCM-LRES) is considerably less potentially predictable. We show that the enhanced potential predictability in the former KCM version can be very likely related to the improved representation of ENSO-like dynamics and its seasonality. We used the statistical models in true forecast mode, i.e. the prediction schemes were trained from data independent of the forecast period. Using observed SSTAs to train the LIM yields significant skill in forecasting observed SSTAs at lead times of up to 4 months across all calendar months, which is mostly restricted to the northern and equatorial western Tropical Atlantic. Similar patterns, but with lower skill, are found when the models’ SSTAs are used, in which LIM trained with the KCM-HRES generally yields higher skills than that from the KCM-LRES. Applying AF yields significant skills in predicting observed SSTAs over the same regions, but the forecast skills are considerably smaller. When the SSTAs together with either sea level pressure (SLP) anomalies or dynamic sea level (DSL) anomalies from the KCM are used to construct the statistical models, the prediction of observed equatorial Atlantic SSTAs can be improved, with significant skill enhancement at lead times of up to 4 months in limited regions. An optimal initial SSTA pattern is found, which results in the largest transient anomaly growth over the entire domain. Independent of external forces, this amplification is developed internally; meaning that the seasonal forecast might be more sensitive to initial conditions than currently thought.