Super-Resolution Remote Sensing Image Classification Based on Residual Networks

This paper presents a deep learning-based superresolution reconstruction method for remote sensing images, targeting the shortage of high-resolution datasets in classification tasks. To address the challenge of using lowresolution images for high-resolution classification, a superresolution generator based on residual networks is developed. This model reconstructs low-resolution remote sensing images into high-resolution ones, effectively expanding the dataset. The study focuses on five types of military aircraft (B1, B52, C17, C130, and KC 135). The reconstruction performance is evaluated using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and reconstruction accuracy. Results show that the reconstruction accuracy reaches 93.5%, while the classification accuracy increases from 76.3% for lowresolution images to 89.1% after super-resolution processing. This enhancement significantly boosts the classification model's accuracy and generalization ability. Additionally, sensitivity analysis of key parameters in the reconstruction process is conducted, offering practical insights. The proposed method effectively leverages historical remote sensing data, resolving the bottleneck of insufficient datasets for classification tasks.