A Deep Learning Method for Radiometric Harmonization of Non-Overlapping Remote Sensing Images

Conventional relative radiometric normalization (RRN) methods establish mapping relationships by using overlapping areas between images to achieve radiometric alignment between them. However, these methods become inapplicable when stitching weakly overlapping or non-overlapping images. We propose a novel radiometric harmonization method that addresses radiometric alignment as a style transfer problem using the CycleGAN, a Generative Adversarial Network architecture. We use two non-overlapping image sets to train the model, and the trained model can perform style transfer between the target image set and the non-overlapping image set. The corrected image closely approximates the conventional RRN result using the real reference image (overlapping with the target image), and is significantly better than the conventional RRN results obtained using the non-overlapping pseudo-reference image.