A Noise-Robust Intelligent Change Detection Framework via Deep Feature Restoration and Posterior Probability Modeling

Change detection under Gaussian noise is challenging because noise perturbs spectral clustering and posterior inference. This study presents a scene-adaptive weakly supervised framework that combines self-supervised single-image restoration with posterior-probability change modeling. A channel-spatial attention aggregation network (CAANet) is optimized jointly from the observed bitemporal scene by masked reconstruction without clean-image targets. The observed and restored images are coupled in a restoration-guided fuzzy decomposition, after which a normalized context-sensitive Bayesian network converts soft signal evidence into land-cover posterior vectors. Their temporal displacement is measured in a normalized semantic evidence space. Experiments on five remote-sensing datasets use zero-mean Gaussian noise with variance v ∈ {0.01, 0.03, 0.05, 0.07, 0.09}. At v = 0.05, FCC_CAANet achieves OA values of 0.9280–0.9659 and Kappa values of 0.7296–0.9521, obtaining the highest OA and Kappa on all five datasets among the implemented methods. The claims are limited to this controlled synthetic Gaussian-noise setting.