SAR polarimetry-based machine learning method for flood detection in semi-urban vegetation systems

Semi-urban vegetation systems play a critical role in ecosystem stability but are increasingly exposed to flood hazards due to climate variability and rapid land-use change. Accurate flood detection in such system remains challenging because radar backscatter is influenced by complex and mixed scattering mechanisms arising from vegetation, built-up structures, and surface water. Conventional intensity-based flood indices struggle to separate flooded vegetation from non-flooded rough surfaces and tend to miss inundated areas under mixed land-cover conditions. To address these limitations, this study presents a physically interpretable flood detection framework that integrates Synthetic Aperture Radar polarimetric descriptors with a machine learning classifier. The proposed approach utilizes dual-polarized Sentinel-1 SAR data to derive polarimetric features from Stokes parameters and the covariance matrix. Specifically, the Degree of Polarization and Linear Polarization Ratio are combined with eigenvalue-based information to capture changes in both amplitude and polarization state between pre-flood and during-flood conditions. These descriptors are integrated into a novel Flood Index (FI) designed to distinguish flooded urban areas dominated by double-bounce scattering from flooded vegetation characterized by depolarized volume scattering. Unlike commonly used indices such as the Normalized Difference Flood Index (NDFI) or VH/VV ratio, the proposed FI exploits polarization behaviour rather than relying solely on backscatter intensity. A Random Forest classifier is trained on the proposed FI using a tile-based sampling strategy to handle class imbalance between flooded and non-flooded pixels. The framework is evaluated across three flood events representing diverse geographic and land-cover conditions: the 2019 Typhoon Hagibis flood in Japan, the 2023 Yamuna River flood in India, and the 2023 Larissa flood in Greece. Model performance is assessed using multiple accuracy metrics, including F1 score, Intersection over Union (IoU), False Positive Rate (FPR), and False Negative Rate (FNR). Results demonstrate that the Random Forest model trained on the proposed Flood Index consistently outperforms threshold-based Otsu methods and NDFI across all study areas. The approach achieves F1 scores ranging from 0.81 to 0.86 and IoU values between 0.70 and 0.76, while maintaining a relatively low False Negative Rate (0.09-0.17), that is critical for minimizing missed flooded areas in disaster response applications. Sensitivity and ablation analyses further confirm the robustness of the Flood Index to speckle noise and highlight the complementary contribution of its individual components. Overall, the proposed framework offers a transferable and computationally efficient solution for flood mapping in semi-urban vegetation systems using widely available dual-polarized SAR data. The results highlight its potential for scalable flood monitoring and rapid damage assessment across regions with heterogeneous land-cover conditions.