Frequency-Decoupled Enhancement and Large Kernel Cross-Scale Fusion Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection (CD) faces three major challenges: pseudochanges induced by illumination and seasonal variations, blurred boundaries caused by insufficient high-frequency constraints, and drastic scale variations across changed objects. Since existing methods rely predominantly on spatial-domain modeling, in which genuine structural changes are inherently entangled with edges, background noise, and textures, these challenges can hardly be resolved within a single framework. To this end, we propose a frequency-decoupled enhancement and large kernel cross-scale fusion network (FDLNet), which jointly exploits spatial and frequency representations to disentangle genuine changes from interference. Specifically, a frequency-enhanced change feature extraction module is designed, in which a frequency feature decomposition module decomposes bitemporal features into complementary frequency components, thereby alleviating feature entanglement and suppressing pseudochanges in cluttered backgrounds. A feature refinement and aggregation module is further proposed to adaptively modulate the decomposed components through cross-attention interaction, refining the representations of genuine structural changes. Furthermore, a large kernel cross-scale fusion decoder is constructed, where 3-D convolutions model interscale dependencies along an ordered scale axis to accommodate drastic scale variations, and large kernel decomposition enables precise boundary restoration. Extensive experiments on three public datasets demonstrate that FDLNet achieves state-of-the-art performance, ranking first in both F1-score and intersection over union on all three datasets with a favorable accuracy–efficiency tradeoff, and exhibits pronounced advantages in scale-varying and boundary-sensitive scenarios. The source code will be available online.