research / PROJECT NOTE / 2026
BharatFloodSAR: turning Sentinel-1 imagery into a flood benchmark
20 flood events. 12 states. A bitemporal SAR benchmark for India.
285 GBof Sentinel-1 GRD imagery processed
Python · PyTorch · Earth Engine · QGIS · W&B
BharatFloodSAR-v1 is an India-specific bitemporal SAR benchmark curated across 20 flood events from 2015–2024, 12 states and 7 terrain types. The data pipeline processed 285 GB of Sentinel-1 GRD imagery into approximately 2,400 annotated 256×256 chips.
The workflow combined Otsu thresholding and JRC water masking with QGIS validation against NRSC Bhuvan maps, achieving at least 80% spatial agreement.
For model pretraining, I worked with UNet++ using EfficientNet-B4, Swin-Tiny UNet and SegFormer-B0 on the Modified Sen1Floods11 dataset, with training runs tracked in Weights & Biases.
[Explore my GitHub]Project overview adapted from Shishir Mishra’s supplied resume.
