Optimal feature-based InSAR phase filtering framework using convolutional neural network and mathematical morphology

Denoising interferometric phase data is a critical processing step for precise topographic reconstruction and land deformation analysis. A noisy interferogram leads to phase unwrapping challenges and may ultimately degrade the quality of various InSAR final products. In this paper, an optimal feature-based interferometric phase image filtering framework is proposed. The filtering process begins by extracting a set of features from simulated interferogram phase image that captures the primary structural elements of the interferometric phase data. Then, a set of soft morphological filters is optimized using the interferogram features set in the complex number field. A convolutional neural network is developed and employed to match patches of real interferogram phase images with the set of interferogram phase image features. The interferogram phase image patches are then filtered with their corresponding optimal filters. The convolutional neural network is trained using a dataset of interferogram phase image features with various noise variance. The proposed framework considers the effect of interferogram noise level on filtering performance by optimizing the filter parameters according to the level of the interferometric phase noise. Experimental results on simulated and real interferogram phase images confirm that the proposed interferogram filtering framework achieves efficient interferogram phase unwrapping as it contributed to high equivalent number of looks (ENL) of 5921 and a minimal number of phase residues (NOR) of 123. The results also approve that the proposed approach has perfect preservation of fringe pattern and interferogram fine detail compared to other state of the art approaches.