Two-Stage Deep Learning Framework For Satellite Image Classification

Abstract

This paper presents a novel two-stage deep learning framework for high-resolution satellite image classification, combining Enhanced Super-Resolution Generative Adversarial Networks with InceptionV3-based transfer learning. Our approach addresses the critical challenge of low-resolution input imagery by first applying a modified ESRGAN architecture with Residual-in-Residual Dense Blocks to perform 4× super-resolution (128×128 to 512×512 pixels), achieving significant improvements in image quality (28.4 dB PSNR, 0.87 SSIM) while maintaining real-time processing speeds (18.2 ms/image). The enhanced images are then classified through a fine-tuned InceptionV3 model, demonstrating superior performance across seven land cover categories (agriculture, airplane, buildings, forest, golf course, river, and tennis court). Experimental results show a 14% average increase in F1-score compared to direct low-resolution classification, with particularly dramatic improvements for small objects (airplanes: +15%) and geometrically complex classes (golf/tennis courts: +19%). The complete system operates at 55 FPS on an NVIDIA A100 GPU, proving its practical viability for real-time satellite image analysis. This work establishes that super-resolution pre-processing can substantially boost classification accuracy of 95% without compromising deployment efficiency, especially for challenging fine-grained categories in remote sensing applications.