Towards Operational Deep-learning Based Damage Proxy Mapping using Synthetic Aperture Radar

Kurzfassung

Damage Proxy Mapping (DPM) is a powerful remote sensing tool to effectively deploy emergency resources (technical, personal, or financial) in the aftermath of natural disasters. Organizations such as the United Nations Satellite Center (UNOSAT) or the European Union s Copernicus Emergency Management Service (Copernicus EMS) produce such damage maps by comparing pre- and post-event optical and radar imagery acquired from air- or spaceborne sensors. The maps are then distributed to governments and first responders of the affected regions. Due to the urgent nature of the product, the focus in generating DPMs is on speed and reliability, using the often-limited datasets at hand rather than producing the highest-quality output, which could take months or even years to generate. Typically, DPMs are created through manual mapping from high resolution optical imagery, which offers easy visual interpretation. However, machine learning approaches are progressively being adopted to further automate and significantly accelerate production and response workflows. A key limitation of optical imagery, however, is its reliance on clear skies. To address this, methods based on Synthetic Aperture Radar (SAR) have been proposed as SAR provides an all-weather and day-and-night data acquisition capability. Yet, SAR-based damage analysis is hindered by its lower resolution (at least regarding spaceborne sensors), higher computational demands, and the complexity of interpreting radar data. In recent years, deep-learning based DPMs have shown great potential. These methods handle the large volume of radar datasets effectively and produce interpretable results rapidly once trained. In particular, Stephenson et al. (2022) used a Recurrent Neural Network (RNN) on transformed coherence timeseries to produce DPMs for various study regions affected by earthquakes. A key factor for their successful performance was the use of the entire coherence timeseries available up until the event date, which could span several years. However, this required the preprocessing of large data collections for each single study area, including the coregistration of stacks of single look complex (SLC) imagery and the computation of a coherence timeseries in the highest possible spatial and temporal resolution. The significant storage and computational demands of these steps likely contributed to the method never being widely adopted by operational DPM providers (to the knowledge of the authors). In this study, we present initial steps to adapt the Stephenson et al. (2022) method to the terrabyte High Performance Computing (HPC) cluster, a cooperation by the Leibniz Supercomputing Center (LRZ) and the German Aerospace Center (DLR). The goal of terrabyte is to enable both operational Earth observation data processing as well as research and development. This infrastructure is also used by DLR s Center for Satellite Based Crisis Information (ZKI), which produces DPMs and other disaster-related products from SAR imagery (though not yet based on RNN-techniques). Our study aims to facilitate future integration of an RNN-based workflow into cluster-based operational disaster response processors. Among the improvements we present are data-parallel training, the integration of coherence timeseries generation into operational SAR processors, the preprocessing of high seismic risk areas in preparation of potential future earthquake events, and the incorporation of ancillary datasets (e.g., precipitation, topography, and land use) into the training algorithm to assess their effect on the assessment accuracy.