Deep Learning-Based Flood Extent Mapping and Depth Estimation Using EOS-04 C-Band SAR: A Case Study of the 2023 Krishna River Basin Flood, India

Introduction: Flooding is among the most frequent and damaging natural hazards, making rapid, reliable mapping of flood extent and depth essential for disaster response. However, persistent cloud cover during active monsoon events severely limits optical remote sensing, and automated SAR-based pipelines that jointly deliver extent and depth remain limited. This study addresses this gap with an end-to-end deep learning pipeline for flood extent detection and depth estimation using dual-polarisation C-band synthetic aperture radar (SAR) imagery from EOS-04 (RISAT-1A) over the Krishna River Basin, Andhra Pradesh, India, during the July 2023 flood event.

Methods: Pre-flood (20 January 2023) and peak-flood (26 July 2023) acquisitions in HH and HV polarisations were calibrated to sigma-naught (σ⁰) and processed with Refined Lee speckle filtering. Seven features were derived: σ⁰_HH, σ⁰_HV, cross-polarisation ratio (σ⁰_HH/σ⁰_HV), Radar Forest Degradation Index (RFDI), Δσ⁰_{_}HH, Δσ⁰_{_}HV, and Gray-Level Co-occurrence Matrix (GLCM) texture. A U-Net segmentation model was trained on 512×512 patches using Otsu-derived pseudo-labels with a combined Dice and binary cross-entropy loss. Inference used sliding-window tiling across the 5152×6928 pixel scene. Flood depth was estimated via the SAR Topographic Analysis (STA) technique, integrating backscatter intensity, temporal change magnitude, and Copernicus GLO-30 DEM elevation data.

Results: The U-Net achieved an Intersection over Union (IoU) of 0.794, F1-score of 0.885, and Cohen's Kappa of 0.874, outperforming the Otsu baseline (IoU ≈ 0.65). Inundation extent reached 102,658 ha across 2,293 flood polygons. STA-derived depths ranged from 0.1–6.0 m (mean: 3.07 m), with peak distribution at 2.5–3 m (18.5% of the inundated area).

Conclusions: This work demonstrates that EOS-04 dual-polarisation SAR, combined with deep learning segmentation and STA-based depth estimation, provides a transferable framework for near-real-time operational flood mapping. Future work includes ground-truth validation and extension to multi-event, multi-basin scenarios.