Benchmarking lightweight deep networks for real-time flood segmentation in unmanned aerial vehicle imagery

The ability to generate up-to-the-minute flood maps using remote sensing imagery is crucial for efficient environmental monitoring and catastrophe response. In order to quickly segregate floods from aerial and Unmanned Aerial Vehicle (UAV) images, this paper provides a comprehensive evaluation of two lightweight convolutional neural network architectures, namely Fast-SCNN and BiSeNetV2. Using a publically accessible flood segmentation dataset, I establish an end-to-end pipeline and evaluate the two models for accuracy (Intersection over Union (IoU), Dice coefficient, F1-score) and speed (inference time per picture). With a validation set F1-score of 0.738 and a peak IoU of 0.608, extensive trials demonstrate that Fast-SCNN outperforms all other segmentation methods. Although BiSeNetV2's inference time is 0.27 ms per picture, it achieves comparable results. Although both algorithms are good at detecting flooded areas, qualitative research shows that they are not very good at defining complex borders or small flood patches. Findings highlight benefits and drawbacks of operational flood monitoring using state-of-the-art real-time semantic segmentation networks. This provides helpful details for implementing remote sensing-based automated flood detection systems.