Research on Remote Sensing Image Classification Based on Transfer Learning and Data Augmentation

Traditional algorithms are no longer effective in the context of the current proliferation of remote sensing image data and resolution, and the remote sensing image classification algorithm based on convolutional neural net-work architecture needs a significant amount of annotated datasets, and the creation of these training data is labor-intensive and time-consuming. Therefore, using a small sample dataset and a mix of transfer learning and data augmentation, this paper suggests a method for classifying remote sensing images. In this paper, the parameters from the Resnet50 model's pre-training on the Imagenet dataset are migrated to the Resnet50-TL model and ultimately classified using Log softmax. The NWPU-RESISC45 dataset is used in this study to train the model and for data Augmentation procedures. The experimental findings demonstrate that the ResNet50-TL model performs better than other popular network architectures currently in use. The model can classify objects with an accuracy of 96.11% using only 700 data points per class, resulting in a high accuracy rate in a limited amount of data. In the future, the dataset will be increased and the network architecture will be updated frequently to make remote sensing picture interpretation more intelligent and portable.