DeepSARFlood is a deep learning-based tool for rapid and automated flood inundation mapping using SAR data. By leveraging weak flood labels from concurrent optical imagery, our approach refines flood detection using cutting-edge Convolutional Neural Networks (CNNs) and Vision Transformer (ViT) architectures. Optimized through multitask learning and model soups, DeepSARFlood integrates a novel gain algorithm to enhance ensemble diversity and provide robust uncertainty quantification. Our results demonstrate that ViT-based and CNN-ViT hybrid models surpass traditional CNNs, achieving a state-of-the-art IoU score of 0.72 on the Sen1Floods11 dataset, setting a new benchmark for flood mapping accuracy and reliability.
This repo provides the codebase for the tool, along with the model weights.
Click to see a video of an example runtime using DeepSARFlood GUI
Sharma, N.K., Saharia, M., 2025. DeepSARFlood: Rapid and Automated SAR-based flood inundation mapping using Vision Transformer-based Deep Ensembles with uncertainty estimates. Sci. Remote Sens. 100203. https://doi.org/10.1016/j.srs.2025.100203
- Sen1Floods11 https://github.com/cloudtostreet/Sen1Floods11
For any queries, please contact:
Nirdesh Sharma
[email protected]
Department of Civil Engineering
Indian Institute of Technology Delhi
Dr. Manabendra Saharia
Principal Investigator, HydroSense Lab
Assistant Professor, Dept. of Civil Engineering
Associate Faculty, Yardi School of Artificial Intelligence
Indian Institute of Technology Delhi
Contact: [email protected]
