Robustness Analysis for VGG-16 Model in Image Classification of Post-Hurricane Buildings
Haoyang Li, Xinyi Wang · 2021
Hurricane is a destructive natural disaster that causes a dramatic loss of life and damage to property. Locating houses damaged by hurricane manually is quite time-consuming and labor intensive. To accelerate the recovery process from the hurricane, AI-based methods such as deep learning networks are widely applied in the field. We hope to use deep learning network models combined with satellite images to quickly classify the houses after the disaster. VGG-16 is one of the most popular used models in the detection of hurricane-damaged houses. To increase the accuracy of model classification, we focus on the robustness analysis of the VGG-16 model on the classification of satellite images of hurricane damage and no-damage houses to see what factors play an important role in the classification decision of the model. In our study, two factors were mainly studied about the influence on the robustness of the model: the distribution of the histogram and the shape of the house. Experiments show that changing one of these two factors will greatly decrease the classification accuracy of the model.