Different CNNs for Enhancing Post-Hurricane Damage Classification

Abstract

Over the past decades, post-hurricane measurements are usually done by humans driving around the affected region to note down manually, which is mostly correct but time-consuming. In this case, more effective methods with CNN models could be proposed since they are advanced at processing and classifying image data. This paper proposed several different CNN models to tackle this issue, which can be utilized to enhance post-hurricane damage measurement by classifying images of buildings from the affected region as damaged or not. Firstly, we got the image data from the affected region with labels of whether being damaged. Five models were trained by these image data, which were basic CNN model, CNN model with Batch Normalization, CNN model with Residual Blocks, pre-trained VGG-19 with frozen weights, and pre-trained VGG-19 with fine tuning. The experimental results demonstrated that Batch Normalization, Residual Blocks, and Learning Rate Decay could greatly improve the basic CNN model performances. The CNN model with Residual Blocks and Learning Rate Decay can reach 99.65%, which is the best model in our project. Furthermore, we note that fine-tuned pre-trained VGG-19 model can also reach 99.6%. This paper can help the rescuers measure the post-hurricane damage in a faster and more efficient way with our classification models based on CNN.