Probabilistic deep learning methods for capturing uncertainty in SAR-based water segmentation maps

Abstract

The occurrence of hazard events, such as floods, has recognized ecological and socioeconomic consequences for affected communities. Geospatial resources, including satellitebased synthetic aperture radar (SAR) and optical data, have been instrumental in providing time-sensitive information about the extent and impact of these events to support emergency response and hazard management efforts. In effect, finite resources can be better optimized to support the needs of often extensively affected areas. However, the derivation of SAR-based flood information is not without its challenges and inaccurate flood detection can result in non-trivial consequences. Consequently, in addition to segmentation maps, the inclusion of quantified uncertainties as easily interpretable probabilities can further support risk-based decision-making. This study presents the results of two probabilistic convolutional neural networks (CNNs) adapted for SAR-based water segmentation with freely available Sentinel-1 interferometric wide (IW) swath ground range detected (GRD) data. In particular, the performance of a variational inference-based Bayesian convolutional neural network (BCNN) is evaluated against that of a Monte Carlo dropout network (MCDN). MCDN has been more commonly applied as an approximation of Bayesian deep learning. Here, differences in the uncertainties identified in both models are highlighted based on the evaluation of an extended set of performance metrics to diagnose data and model behaviours and to evaluate ensemble outputs at tile- and scene levels. Since the understanding of uncertainty and subsequent derivation of uncertainty information can vary across applications, this study demonstrates how uncertainties derived from ensemble outputs can be integrated into maps as a form of actionable information. The findings highlight how the consideration of both segmentation accuracy and probabilistic performance can build confidence in products used to make informed decisions and to support emergency response within flood situations.