Trs-net tropical revolving storm disasters analysis and classification based on multispectral images using 2-d deep convolutional neural network

Abstract

Damage assessment is one of the most important factors in estimating loss after hurricane. The assessment's findings are very important to identify casualties, temporary housing requirements, financial losses, socio-economic effects, etc. In this paper, our goal is to detect the damage using satellite images captured for the natural disasters like hurricane Harvey. In order to assist the experts to speed up the investigation process for providing the best recommendations this research study presents a compact deep learning-based automatic model to assess hurricane damage using satellite images. TRS-Net (Tropical Revolving Storm Network), model uses flood satellite images to identify the damage caused by the hurricane Harvey. The proposed model effectively extracted features from the satellite images by using depth-wise convolution with varying dilation rates. 23,000 satellite Harvey hurricane images of which 15,000 were damage images caused directly or indirectly and the remaining were the images of normal regions. The model included four clustered layers with convolution and max pooling layer enhanced with batch normalization and dropout to attain better accuracy in detecting the damage caused by hurricane Harvey. The dataset was trained and tested with three different optimizers like Adam, RMSProp and SGD where Adam outperformed the others with 98.57% of testing accuracy. Furthermore, based on the loss and accuracy curve, the performance evaluation of the dataset is analyzed. The experimental findings demonstrate that the suggested TRS-Net model achieves the greatest classification accuracy with 0.99, 0.98, 1.0, 0.96 as F1 Score, precision, sensitivity and specificity respectively. Therefore, the present work assist experts’ in efficient hurricane damage assessment and overcome the resource and time intensive manual assessment method.