Probabilistic SAR-based water segmentation with adapted Bayesian convolutional neural network

Abstract

Geospatial resources, including satellite-based synthetic aperture radar (SAR) and optical data, have been instrumental in providing time-sensitive information about the extent and impact of natural hazard events, such as floods, 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 poor decision-making and non-trivial, adverse consequences. Reliable uncertainty quantification of flood extent estimates addresses this risk. In this context, our 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). In previous studies where uncertainty information has been generated along with segmentation results, MCDN has been more commonly applied as an approximation of Bayesian deep learning. Differences between the two approaches are highlighted with the application of a set of extended performance metrics. Both segmentation and uncertainty outputs are evaluated at data-, model-, tile- and scene-levels. The methods are demonstrated on the binary segmentation of a reference water dataset. We show that while different probabilistic techniques return comparable segmentation accuracies, they are differentiated based on their performance in assigning reliable probabilities. In particular, MCDNs characterized by more restrictive architectures generally lead to overconfident prediction intervals, whereas BCNNs have greater flexibility to learn the mean and the spread of the parameter posterior. Furthermore, we demonstrate that examining the (inaccurate, certain) metric is a better indicator of reliable uncertainty quantification and the BCNN is recommended to quantify uncertainties associated with SAR-based segmentation outputs. This information is especially valuable where the cost of inaccurate detections (false-positive and false-negative) is high. These preliminary findings highlight how the manner in which probabilities are properly assigned and their inclusion are instrumental and complementary to the production of flood masks and should be considered as a standard in the natural hazards domain.