Multi-Source Earth Observation and Explainable GeoAI for Urban Flood Damage and Land-Impact Mapping DATA
Description
Urban flooding increasingly threatens urban land, critical infrastructure, and essential services as climate change, rapid urbanization, and soil sealing intensify hydrometeor-ological risk. Timely post-disaster land and infrastructure mapping is therefore essen-tial for identifying exposed areas, assessing urban land vulnerability, and supporting emergency response, territorial planning, recovery, and resilience strategies. This study presents an integrated framework that fuses Sentinel-1 synthetic aperture radar (SAR) and Sentinel-2 optical data through a hybrid CNN-LSTM architecture enhanced with explainable artificial intelligence (XAI). The framework was evaluated using 20 flood events recorded by the Copernicus Emergency Management Service between 2021 and 2024, representing more than USD 102 billion in economic losses. Fusion of RGB, NIR, and SWIR bands with SAR data achieved an intersection-over-union of 0.7053, corre-sponding to a 20.3% improvement over individual modalities. SAR contribution was strongly associated with the damage index in densely built-up areas, whereas optical information provided greater precision in peri-urban land. SHAP analysis identified Sentinel-2 SWIR bands B11 and B12 as the most influential variables for detecting damage in dense urban environments, followed by VV-polarized SAR backscatter, while NDVI and MNDWI showed greater importance in vegetated peri-urban zones. Grad-CAM maps spatially localized the image regions driving damage classifications, particularly affected buildings, road corridors, and critical-service areas, thereby im-proving the physical interpretability of model predictions. The proposed framework strengthens transparent post-flood damage assessment and provides actionable in-formation for land-use planning, infrastructure recovery, risk reduction, and urban re-silience.
Files
Steps to reproduce
The study applies a quantitative geospatial and computational methodology to assess urban flood damage through multi-source Earth observation and explainable GeoAI. The framework integrates Sentinel-1 SAR, Sentinel-2 optical imagery, deep learning, and explainability techniques. It is validated using 20 flood events documented by the Copernicus Emergency Management Service between 2021 and 2024. The methodology follows six stages. First, Sentinel-1 GRD images in VV and VH polarizations, Sentinel-2 L2A scenes, CEMS flood products, and data on flooded area, affected buildings, infrastructure losses, and event dates are collected. Images are selected for three periods: up to 30 days before the event, 24–72 hours after onset, and 7–30 days afterward. Second, Sentinel-1 data are corrected for orbit, radiometrically calibrated, filtered with a 7 × 7 Lee filter, terrain-corrected using Copernicus DEM, geocoded, and resampled to 10 m. Sentinel-2 data are atmospherically corrected, cloud- and shadow-masked, co-registered with SAR imagery, resampled, and normalized. Third, a 14-feature vector is constructed from VV and VH backscatter, the VV/VH ratio, temporal change in VV, Sentinel-2 bands B2, B3, B4, B8, B11, and B12, and four spectral indices. SAR and optical variables are then integrated through feature-level fusion. Fourth, a hybrid CNN-LSTM model is trained. A ResNet-50 encoder extracts spatial information, while a bidirectional LSTM captures temporal relationships across pre-event, during-event, and post-event images. The model classifies damage into five categories: no damage, minor, moderate, severe, and destroyed. Training uses AdamW, weighted focal loss, dropout, mixed precision, and fixed random seeds. Fifth, SHAP quantifies the global and event-level contribution of each input feature, while Grad-CAM identifies the image regions that most influence each prediction. These outputs verify whether the model focuses on damaged buildings, roads, and critical services. Sixth, performance is evaluated using overall accuracy, weighted F1-score, IoU, Cohen’s kappa, AUC-ROC, and confusion matrices. McNemar’s test, Pearson correlation, fixed-effects panel regression, and K-means clustering support the statistical validation. To reproduce the study, researchers should publish the complete list of CEMS events, satellite scene identifiers, inclusion criteria, damage-index formula, class-recoding rules, selected spectral indices, patch size, data partitions, normalization parameters, hyperparameters, software versions, random seeds, and source code. The workflow should generate GeoTIFF damage maps, class probabilities, event-level metrics, SHAP values, Grad-CAM maps, and the statistical datasets used to produce all tables and figures.