Leveraging power of deep learning for fast and efficient elite pixel selection in time series SAR interferometry

This work proposes an improved convolutional long short-term memory (ConvLSTM) based architecture for selection of elite pixels (i.e., less noisy) in time series interferometric synthetic aperture radar (TS-InSAR). Compared to previous version, the model can process InSAR stacks of variable time steps and select both persistent (PS) and distributed scatterers (DS). We trained the model on ~20,000 training images (interferograms), each of size 100 by 100 pixels, extracted from InSAR time series interferograms containing both artificial features (buildings and infrastructure) and objects of natural environment (vegetation, forests, barren or agricultural land, water bodies). Based on such categorization, we developed two deep learning models, primarily focusing on urban and coastal sites. Training labels were generated from elite pixel selection outputs generated from the wavelet-based InSAR (WabInSAR) software developed by Shirzaei (2013) and improved in Lee and Shirzaei (2023). With 4 urban and 7 coastal sites used for training and validation, the predicted elite pixel selection maps reveal that the proposed models efficiently learn from WabInSAR-generated labels, reaching a validation accuracy of 94%. The models accurately discard pixels affected by geometric and temporal decorrelation while selecting pixels corresponding to urban objects and those with stable phase history unaffected by temporal and geometric decorrelation. The density of pixels in urban areas is comparable to and higher for coastal areas compared to WabInSAR outputs. With significantly reduced time computation (order of minutes) and improved selection of elite pixels, the proposed models can efficiently process long InSAR time series stacks and generate rapid deformation maps.