Katpnet: beyond dense prediction with a two-stage KAN-Transformer for SAR change detection

Synthetic aperture radar (SAR) change detection remains challenging due to severe speckle noise, strong class imbalance, and the widespread use of single-stage dense prediction frameworks that process all patches uniformly, often leading to elevated false-alarm rates. To address these challenges, we propose KATPNet, a two-stage KAN–Transformer architecture for robust and adaptive SAR change detection. In Stage 1, a supervised Proposal Generator employs KAN–Transformer blocks to extract patch embeddings and estimate proposal confidence scores under pseudo-label supervision, enabling the network to distinguish reliable from uncertain regions. Stage 2 introduces a KAN-based decision module that leverages adaptive nonlinearities to model the heavy-tailed statistics of SAR data, while an uncertainty-aware gating mechanism uses the Stage-1 confidence to suppress false alarms. The network is trained end-to-end using weighted binary cross-entropy for proposal supervision and a hybrid Tversky–Cross-Entropy loss to mitigate class imbalance. Experiments on three challenging SAR datasets (Ottawa, Bern, and Yellow River) demonstrate that KATPNet attains competitive or superior performance compared with representative recent methods, reaching a Kappa coefficient of 89.32% on the noise-dominated Yellow River dataset. These results highlight the effectiveness of explicit proposal supervision and adaptive KAN nonlinearities for accurate and generalizable SAR change detection.