Unpaired Satellite Precipitation Downscaling via Multiple Generative Constrains

Satellite precipitation downscaling is designed to enhance low-resolution images to high-resolution images while maximizing the retention of details and ensuring data accuracy. Due to data limitations, supervised learning for satellite precipitation downscaling is challenging, making unsupervised learning mainstream. Current unsupervised downscaling methods for satellite precipitation remote sensing images rely heavily on the automatic detection of fine details and self-correction mechanisms to produce high-resolution images. However, these methods often falter when dealing with highly detailed and complex datasets, which hampers their ability to generate high-accuracy images. To tackle this challenge, this letter proposes a novel unpaired generative adversarial network with multiple generative constraints (MGC-GAN). It alleviates the need for paired training data, which is a common requirement in existing unsupervised methods. Specifically, we integrate multiple generative constraints, including a discriminative loss along with additional perceptual losses, to enhance the quality of the generated high-resolution images. It has been rigorously tested across various datasets, including GSMaP, as well as natural landscape remote sensing image datasets AID and NWPU-RESISC45. Experimental outcomes demonstrate that this innovative method not only significantly improves image resolution but also effectively retains the integrity of the image details. In particular, it achieves 20.345 and 1.41 of PSNR and bias on GSMaP, meanwhile reaching 24.318 and 0.797 of PSNR and SSIM on natural landscape remote sensing datasets, respectively.