Single infrared remote sensing image super-resolution via supervised deep learning

The infrared observation sensors aboard remote sensing satellites play an important role in the applications of crop yield estimation, environmental protection, land resources survey, disaster monitoring, etc. But infrared remote sensing data is always in low resolution due to hardware limitations. It is a high cost-effective choice to improve the spatial resolution of infrared remote sensing data through super-resolution (SR) algorithm. Deep learning methods have made great breakthroughs in super-resolution of natural images. In this paper, we comparably study five recently popular supervised-deep-learning-based single image SR models for the purpose of super-resolving infrared images, including SRGAN, ESRGAN, LapSRN, RCAN, and SRFBN. We first test the performance of models trained by natural images on infrared remote sensing images to obtain a benchmark, and then specially fine-tune the SR models using infrared images of Landsat8 in a transfer-learning manner. We evaluate the performance of all these fine-tuned models on infrared images with three indicators including PSNR, SSIM, and NIQE. The experimental results show that the SRFBN model achieves the best generalization ability and SR performance. Therefore, we suggest using SRFBN for super-resolution reconstruction of single infrared remote sensing image in applications.