High Resolution Remote Sensing Image Classification Based on Deep Transfer Learning and Multi Feature Network

Abstract

To improve the automatic classification accuracy of remote sensing images, this study raises a high-resolution remote sensing image classification model that combines deep transfer learning and multi-feature network. In this paper, deep transfer learning is the core technology of remote sensing image classification model, and VGG16, Inception V3, ResNet50 and MobileNet are used to build a fusion classification model through serial fusion. By testing the fusion model, the Transfer Learning ResNet50-MobileNet (TL-RM) model with the best performance was obtained. Finally, experimental analysis verified its significant stability: the average accuracy of TL-RM on a small sample high-resolution remote sensing image dataset was 96.8%, and the Kappa coefficient was 0.964, both of which were the highest values among all models. The accuracy of this model shows a slight upward trend and then stabilizes as the iterations increases. The training and testing sets accuracy ultimately stabilizes at around 100% and 98%, and the loss value ultimately stabilizes at around 1%. Moreover, TL-RM only has a low classification accuracy for residential areas in remote sensing images, with a classification accuracy of over 97% for other categories. The experiment shows that the TL-RM model has significant accuracy and stability, providing a reliable theoretical and experimental basis for remote sensing image classification research.