A physics-guided neural network for flooding area detection using SAR imagery and local river gauge observations

TL;DR

The study addresses the challenge of mapping flood extent from SAR data in data-sparse river basins by introducing a physics-guided neural network (PGNN) that links flood extent to local water elevations. It uses Sentinel-1 time-series imagery and river-gauge observations, optimizing with a loss based on the Pearson correlation coefficient between the predicted total water extent and observed water levels, denoted as $PCC$. The FCN-8–inspired architecture combines a segmentation head with a regression path and enforces nonnegative weights, enabling unsupervised learning without ground-truth water masks. Across five diverse study sites, the approach achieves higher IoU scores than unsupervised baselines, particularly during low-water conditions, demonstrating potential for regional flood mapping where gauge networks exist, albeit with limitations related to river width, vegetation, ice, and the need for area-specific training or transfer learning.

Abstract

The flooding extent area in a river valley is related to river gauge observations. The higher the water elevation, the larger the flooding area. Due to synthetic aperture radar\textquoteright s (SAR) capabilities to penetrate through clouds, radar images have been commonly used to estimate flooding extent area with various methods, from simple thresholding to deep learning models. In this study, we propose a physics-guided neural network for flooding area detection. Our approach takes as input data the Sentinel 1 time-series images and the water elevations in the river assigned to each image. We apply the Pearson correlation coefficient between the predicted sum of water extent areas and the local water level observations of river water elevations as the loss function. The effectiveness of our method is evaluated in five different study areas by comparing the predicted water maps with reference water maps obtained from digital terrain models and optical satellite images. The highest Intersection over Union (IoU) score achieved by our models was 0.89 for the water class and 0.96 for the non-water class. Additionally, we compared the results with other unsupervised methods. The proposed neural network provided a higher IoU than the other methods, especially for SAR images registered during low water elevation in the river.

Figures (6)

• Figure 1: Architecture of the FPGNN model.
• Figure 2: Visual comparison of water masks obtained by thresholding from various spatial data and the effect of water elevation in the river on the water extent area.
• Figure 3: Average IoU values of water and non-water masks produced by FPGNN models (a) or benchmarking methods (b). Panels are varied by the study sites (columns) and class (rows). Within each panel, the IoU value is shown per threshold used in hard classification or benchmarking method (panel columns) and reference water mask (panel rows).
• Figure 4: Comparisons of IoU for days of optical image acquisition for sites located in Brazil (a), Cambodia (b), Italy (c), England (d) and Poland (e). Panel for each site includes three plots: observed water elevation, calculated water and non-water class IoU for the FPGNN, GMM and SC methods in MNDWI and DTM validation.
• Figure 5: Contingency maps obtained from a pixel-to-pixel comparison between the MNDWI water masks and FPGNN or the benchmarking methods for all study sites for low water elevation. True positives are shown in blue, true negatives in gray, false positives in green, and false negatives in red.