Flood extent mapping using multi-temporal sentinel-2 remote sensing and deep learning

This study evaluates a flood extent mapping framework that combines multi-temporal Sentinel-2 imagery with deep learning in the Karkheh River Basin, Khuzestan Province, Iran. The 2019 flood scenes were processed into fixed-size temporal patches using four reported predictor channels: B3, B8, NDWI, and SMI. Five architectures were evaluated: CNN-LSTM, CNN-LSTM-U-Net, ConvLSTM, Conv-minLSTM, and ConvLSTM-U-Plus. Performance was assessed using intersection over union (IoU), Dice/F1, precision, recall, and binary accuracy. ConvLSTM achieved the highest F1-scores for the balanced 8 × 8 and 16 × 16 configurations and the unbalanced 32 × 32 configuration, whereas ConvLSTM-U-Plus obtained a marginally higher F1-score for the balanced 32 × 32 configuration and produced spatially coherent boundaries in the illustrated examples. The findings support convolutional recurrent architectures for event-based satellite flood mapping, while validation on independent events and regions is required before operational application.