A Model Integrating Matrix Factorization and Adaptive Propagation for Polsar Image Change Detection

Polarimetric SAR (PolSAR) images have become a crucial data source for change detection in cloud cover and fog regions. Deep learning (DL) enables extract change features form PolSAR images automatically, overcoming the limitations of conventional methods. To address the computational complexity of global attention and inaccurate prediction of changes edges in DL, this study proposes a model integrating matrix factorization and adaptive propagation. The model adopts a Siamese structure, receiving bi-temporal PolSAR images as input. Then, a matrix factorization based attention module replaces traditional attention design, utilizing the low-rank matrix from matrix factorization as global information. Additionally, an adaptive propagation module is introduced to improve edge pixel sampling, calculating central pixel values based on fixed and variable offsets from a local neighborhood. Experimental results on Sentinel-1 dual-pol SAR dataset demonstrate the model's effectiveness in preserving edge integrity and enhancing land change detection accuracy.