Transfer Learning Performance Analysis for VGG16 in Hurricane Damage Building Classification

While hurricane causes great damage in resident buildings, very few efficient responses can be delivered to rescuers. Combining satellite image and Convolutional Neural Network (CNN) transfer learning, rescuers can locate the damaged buildings in time. Therefore, it is crucial to determine the factors of transfer learning performance in this case. In this paper, we used VGG16 as our base model. We investigated it from three aspects: 1. the input image size, 2. the network structure (including adding filters and changing top dense layers), 3. the classifier activation function. The results show the size of the input images influences the performance the most, and the activation function of the classifier has the smallest effect.