Remote Sensing Image Classification Based on Multi-Spectral Cross-Sensor Super-Resolution Combined With Texture Features: A Case Study in the Liaohe Planting Area

Abstract

High-resolution (HR) optical remote sensing images are typically small in swath and, due to cloud cover, their revisit period, mosaic error, and other problems, it is often infeasible to obtain a large range of remote sensing images for a study area. Meanwhile, low-resolution (LR) satellite images suffer from insufficient spatial and texture information for ground objects. Therefore, classifying a study area with high spatial resolution, large area, and no cloud occlusion using optical remote sensing imagery is very difficult. In recent years, the rapid development of super-resolution reconstruction (SRR) technology has made high-quality spatial resolution reconstruction possible. The SRR of real images is usually accompanied by problems such as sensor spectral range differences, cloud occlusion in the research area, and the SRR algorithm sacrificing a lot of the original information. In this study, with an improved PGGAN, we use only a small number of samples, the wide-swath medium-resolution satellite was restored to the same resolution as the high-resolution satellite, a new method for SRR multi-spectral optical remote sensing image classification based on texture reconstruction information is proposed, and a wide range of high-precision feature classifications are achieved in the study area. In order to solve the problem of spectral distortion in the process of multi-spectral image SRR and the weak generalization of optical remote sensing image ground object classification due to the difference between temporal and spatial features, we combine the idea of ground object classification with texture features obtained after super-resolution reconstruction. We used the support vector machine (SVM) and random forest (RF) classification methods to evaluate the classification effect of each texture spectral feature combination, with the overall accuracy (OA) of the SVM and RF classifiers reaching 98.93% and 98.51%, respectively. The land-use and land-cover (LULC) classification accuracy of the SRR images combined with texture features is much higher than that when directly classifying the original GaoFen-1 and Sentinel-2 images. The obtained results imply that the method of superimposing texture features allows for better classification results in the Liaohe estuary area, providing a new technical idea for the study of LULC classification.

Keywords

• Super-resolution
• deep learning
• generative adversarial network
• land use and land cover
• multi-spectral remote sensing