Remote sensing image scene classification based on ResNeSt

Abstract

The study used the deep learning method to achieve the natural scene classification of remote sensing images, which were taken by the satellite Tiangong-2. Because of the diversity of remote sensing images, a Convolutional Neural Network (CNN) model that can complete the task of classifying natural scenes of remote sensing images was constructed using the variant ResNeSt based on the Residual Neural Network (ResNet). The NaSC-TG2 remote sensing image dataset released by the Space Application Engineering and Technology Center of the Chinese Academy of Sciences was used in this work. The dataset consists of 20,000 photos that are grouped into ten scene groups on average, with 2,000 images per scene category. And nine models including ResNet50, ResNet101, ResNet200, SE-ResNet50, SE-ResNeXt50, SE-ResNeXt101, SE-ResNeXt152, ResNeSt50, ResNeSt101 and ResNeSt200 were compared and tested on the NaSC-TG2 dataset. After training and testing on the dataset, ResNeSt101 achieved better results than other comparative models in the end, with the highest accuracy of 98.52% on the testing sets. This study offers a technique for categorizing remote sensing picture scenes and has made some significant contributions to space geoscience and application research.