Does High-Frequency Matter? A Remote Sensing Change Detection Network with High-Frequency Selection and High-Order Recursion

Remote Sensing Change Detection (RSCD) benefits from frequency domain analysis. However, existing methods are often confined to first-order decomposition and tend to amplify high-frequency signals indiscriminately, which can exacerbate noise and compromise detection performance. We reveal the crucial role of high-frequency components in capturing boundaries and fine details, supported by compelling visualizations and experiments that substantiate this key contribution. Motivated by this, we propose a RSCD Network with High-frequency Selection and high-Order Recursion (High-SORNet), where the core idea is to achieve robust structural extraction through high-order recursion and multi-level interaction exclusively in the low-frequency domain, significantly enhancing feature stability and semantic coherence. To further enhance feature discrimination, we introduce High-Frequency Selection Triple Constraint (HFSTC) Loss that enforces sparsity, cross-scale consistency, and gradient alignment, effectively enhancing effective details while suppressing noise and irrelevant textures. Experimentally, our High-SORNet consistently outperforms state-of-the-art methods, achieving IoUs of 54.80% (CLCD), 88.91% (GoogleBuilding), and 83.59% (PRELEVER).