AFDINet: Adaptive frequency disentanglement and interaction in hybrid Mamba–CNN for optical-SAR land cover classification

Abstract

The synergistic integration of optical and Synthetic Aperture Radar (SAR) imagery enables robust Earth observation by overcoming the limitations of single-sensor modalities, such as atmospheric occlusion and spectral deficiency. However, prevailing fusion paradigms typically treat multi-modal integration as a homogeneous spatial operation, neglecting the distinct physical properties inherent in different frequency domains. This oversight often results in the loss of high-frequency (HF) details, leading to fragmented boundaries in fine-grained land cover classification (LCC). To bridge this gap, this study introduces the Adaptive Frequency Disentanglement and Interaction Network (AFDINet), a hybrid architecture fusing CNN and Mamba. We devise a Learnable Frequency Disentanglement (LFD) module that reformulates wavelet decomposition as a data-driven process, adaptively separating structural consensus from modality-specific textures. Subsequently, a High–Low Frequency Fusion (HLF${}^3$) module implements a role-aware interaction strategy, calibrating low-frequency (LF) geometry while enhancing HF details. Extensive empirical evaluations on the Xi’an, YYX-OPT-SAR, and WHU-OPT-SAR datasets demonstrate that AFDINet achieves state-of-the-art (SOTA) performance in terms of OA, mIoU, and Kappa, while exhibiting exceptional resilience to cloud interference.