Comparative Analysis of Despeckling Filters Based on Generative Artificial Intelligence Trained with Actual Synthetic Aperture Radar Imagery

Abstract

The speckle is a granular undesired pattern present in Synthetic-Aperture Radar (SAR) imagery. Despeckling has been an active field of research during the last decades, with approaches from local filters to non-local filters that calculate the new value of a pixel according to characteristics of other pixels that are not close, the more advanced paradigms based on deep learning, and the newer based on generative artificial intelligence. For the latter, it is necessary to have a large enough labeled dataset for training and validation. In this study, we propose using a dataset designed entirely from actual SAR imagery, calculated by multitemporal fusion operations to generate a ground truth reference, which will yield the models to be trained with the actual speckle patterns in the noisy images. Then, a comparative analysis of the impacts of including the generative capacity in the models is performed through visual and quantitative assessment. From the findings, it is concluded that the use of generative artificial intelligence with actual speckle exhibits notable efficiency compared to other approaches, which makes this a promising path for research in the context of SAR imagery.