The Study of Transferable Learning in Satellite Images Classification with Different Convolution Neural Networks

In the past, the detection of damaged buildings based on satellite images after hurricane is labor-intensive. The recent development of Convolutional Neural Network (CNN) makes possible a more effective method for the classification task. This paper seeks to compare the accuracy of different convolutional neural network architectures such as ResNet, VGG, MobileNet, and InceptionV3 with and without transfer learning. With Hyperparameter Configuration, a set of hyperparameters are adjusted and utilized for each architecture to achieve a highest accuracy. After the model construction, optimization, and architecture comparison for both with and without transfer learning, ResNet model with transfer learning demonstrates the best performance, with an accuracy of 99.25% on the test dataset. Moreover, this paper aims to investigate into the nature of transfer learning on different architectures when implementing classification task with satellite images and explore how the model structure influence the training of classifying satellite images with and without Transfer Learning. By exploring the model structures and visualizing the feature maps of the representative models such as ResNet and VGG both with and without transfer learning, this paper finds that the suitability of transfer learning in satellite images may be affected by the characteristics of the model architecture such as the residual block of ResNet model and that the brightness of the feature map may influence the accuracy of the model, which is related to whether transfer learning is implemented.