Damage analysis on post-hurricane images taken by satellite based on multiple variants of convolutional neural networks
Jiawei Guo, Shixuan Wu · 2021
Due to the difficulty in locating the unaccounted individuals for or injured in captured satellite pictures, the post-hurricane rescue team wastes a bountiful effort in searching. Thus, it is out of necessity to propose a more effective method that can be used to analyze the related images. In our study, we proposed a stacked convolutional neural network architecture. We employed different classical models, e.g., VGG-16 and MobileNet, to detect the damaged or non-damaged buildings impacted by a hurricane based on satellite imagery. To be more specific, we implemented just 6 convolutional layers to achieve a result that can be almost as well as VGG-16. At the very beginning, we implemented data augmentation on our dataset, which can enhance the stability and generalization ability of the model. What is more, we noticed that the images are collected by satellite which lacks details. Thus, we can use fewer layers that can boost the computational time and ensure the model's accuracy, so we balanced the tradeoff between model complexity and model accuracy. Moreover, in our study, we set VGG-16 and MobileNet as our baseline. Therefore, the difference in performances between each model can be seen clearly. Conclusively, our proposed model has roughly the same performances regarding classification accuracy and losses on the best transfer learning model VGG16 and other models. The processing time of our model is less than the other models due to the simple architecture.