Explainable Flood Segmentation on Sentinel-1 SAR Imagery: A Comparative Study of CNN and Transformer Architectures

2026-06-15 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Environmental Science & Earth Systems · Depth: Expert, quick

Summary

A comprehensive study compared convolutional neural network (CNN) and vision transformer architectures for multi-class flood segmentation using Sentinel-1 Synthetic Aperture Radar (SAR) imagery. The research focused on distinguishing flooded land from permanent water bodies and land, crucial for disaster response. Three CNN models (U-Net, U-Net++, DeepLabV3 with ResNet-34) and three SegFormer variants (b0, b1, b2) were evaluated on the ETCI NASA dataset and SenFloods11, employing scene-based data splits for realistic spatial generalization assessment. SegFormer-b2 significantly outperformed the U-Net baseline on the ETCI dataset, showing higher flood IoU across all 7 test scenes. While this advantage narrowed on Sen1Floods11 after fine-tuning, it remained notable for spatially fragmented flood events. The study also incorporated explainability techniques, revealing SegFormer-b2's more spatially coherent Grad-CAM activations and U-Net's superior uncertainty estimates along flood boundaries.

Key takeaway

For Machine Learning Engineers developing flood segmentation systems, this study suggests considering SegFormer-b2 as a primary architecture, especially for its strong performance on SAR imagery. If your application requires precise flood boundary uncertainty estimates, U-Net might be a more suitable choice. Always validate models using scene-based data splits to ensure robust spatial generalization in real-world disaster response scenarios.

Key insights

SegFormer-b2 excels in SAR flood segmentation, outperforming CNNs on specific datasets, with distinct explainability characteristics.

Principles

• SAR sensors enable all-weather flood monitoring.
• Scene-based data splits validate spatial generalization.
• Explainability reveals model decision nuances.

Method

Compared U-Net, U-Net++, DeepLabV3, and SegFormer variants for multi-class flood segmentation on Sentinel-1 SAR imagery. Evaluated on ETCI and SenFloods11 datasets using scene-based splits, incorporating qualitative and quantitative explainability.

In practice

• Deploy SegFormer-b2 for robust flood segmentation.
• Prioritize U-Net for boundary uncertainty estimates.
• Use scene-based splits for realistic evaluation.

Topics

• Flood Segmentation
• Sentinel-1 SAR
• Vision Transformers
• Convolutional Neural Networks
• Explainable AI
• SegFormer

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• LALE: Lightweight-Transformer Architecture for Land-Cover Estimation — LALE achieves efficient remote sensing image segmentation by combining resolution-specific encoders and an all-MLP decoder.
• Beyond Backscatter: InSAR coherence from detected SAR images — A Residual U-Net regresses InSAR coherence directly from detected SAR images, bypassing coregistration and improving accuracy for large-scale applications.
• Unlocking Diffusion Hierarchies: Adaptive Timestep Selection for Zero-Shot Segmentation — Diffusion models exhibit hierarchical semantic progression, enabling adaptive timestep selection for improved zero-shot segmentation.
• U-Net Explained: The CNN That Revolutionized Image Segmentation — U-Net's U-shaped encoder-decoder with skip connections enables precise pixel-level segmentation from limited data.
• FRED: A Multi-Modal Autonomous Driving Dataset for Flooded Road Environments — FRED is the first multi-modal dataset for autonomous driving in flooded environments, enabling robust water hazard detection.
• GMBFormer: An NDVI-Guided Global Memory Bank Transformer for Urban Green-Space Extraction from Ultra-High-Resolution Imagery — GMBFormer improves urban green-space extraction by decoupling NDVI and using a global memory bank for semantic reuse.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.