sidgan/ETCI-2021-Competition-on-Flood-Detection

Implements semi-supervised segmentation using cyclical pseudo-labeling and noisy student training to leverage unlabeled SAR data, achieving ~0.76 IoU on flood detection tasks. Built on PyTorch with segmentation_models and DenseCRF, optimized for distributed training via DistributedDataParallel and mixed-precision (torch.cuda.amp). Targets geospatial disaster response workflows, with inference accelerated to 1.22 seconds per 63,000 km² tile and operationally deployed across multiple countries for early warning systems.

179 stars. No commits in the last 6 months.