Natural Disaster Prediction from Sentinel-2 Images using Convolutional Neural Network

Abstract

Flooding and other natural disasters may have a devastating impact on property and human life. We need a precise flooding evaluation of the impacted region after the occurrence to avail rescuing through a crisis reaction unit. Obtaining an accurate estimate of a flooded region traditionally requires a huge amount of human resources. To overcome this limitation, we proposed an automatic water flood prediction technique with the combination of water indexes to predict natural disasters. The proposed model is evaluated on the Sentinel-2 satellite imagery. With VGG16, we investigated several water indexing approaches and developed a water detection function based on the Green/SWIR and Blue/NIR bands. Our investigation indicates that when coupled with the VGG16 network, our strategy outperforms all existing water index techniques in detecting natural disasters from Sentinel-2 imagery. The proposed approach shows a better performance out of all other techniques with an accuracy value of 0.953%.