Self-supervised learning of deep despeckling networks with MERLIN: ensuring the statistical independence of the real and imaginary parts

Abstract:
Due to the wide variety of sensors, with different spatial resolutions, operating frequency bands, as well as acquisition modes (Stripmap, Spotlight, TOPS...), despeckling neural networks trained on a given type of SAR images do not generalize well. By directly training on images from the sensor and acquisition mode of interest, self-supervised learning is a very appealing solution. This paper analyses the preprocessing requirements of the MERLIN strategy that assumes statistical independence of the real and imaginary parts of single-look-complex SAR images to perform the self-supervised training. Adequate spectral corrections are proposed to handle asymmetrical spectra and moving Doppler centroids.