Reduction deep learning model for floods recognition in satellite images

Satellite images to forecast natural disasters are being researched and improved. This paper proposes a deep learning model that uses a combination of RGB and nIR satellite images to identify flood regions. The proposed model has lower computation complexity (42.51GMac), reduces 78.67% compared with U-NET. In this research, a new loss function focusing on water/flood areas is proposed, with the average AP and IoU values over the flood zones being 87.7% and 79.5%, respectively.