Enhancing Remote Sensing Image Classification by Integrating Transfer Learning and Random Forests

Aerial ground images contain multi-level patterns and features, and the classification accuracy of the model for optical remote sensing scene images is limited due to the complex spatial patterns, similarity between classes, and high-class diversity. In this paper, an optical remote sensing scene classification algorithm based on random forests and transfer learning is proposed. Firstly, multi-level feature information is extracted from the MobileNetV2 model, Vgg19, and then the features are combined. The features extracted from both algorithms are filtered based on the importance of the feature to the random forest model chosen for classification, and finally the model is trained on the final features. The proposed method is verified by experiments on a public remote sensing dataset, UCM. Compared with other advanced scene classification methods, this method achieves better classification performance, achieving an accuracy of 97.14%.