Multimodal Data Fusion Classification via Adaptive Frequency Domain Sparse Enhancement

Achieving efficient fusion of multimodal remote sensing (RS) data from spaceborne platforms for improved land cover classification performance remains a major challenge in remote sensing research. However, existing multimodal data fusion methods in the frequency domain are often affected by noise interference, feature redundancy, and feature imbalance. To address these issues, this article proposes an adaptive frequency domain sparse enhancement (AFDSE) architecture, designed to improve classification performance through frequency feature enhancement and joint multimodal feature learning. Specifically, an adaptive frequency enhancement (AFE) module is designed, which combines task-driven dynamic weight learning with Laplacian convolution kernels to selectively enhance local detail information. In addition, a sparse enhancement (SE) module is introduced, which first captures spatial correlations using a sparse attention mechanism (AM), then enhances global structural information through a feature enhancement network, and finally applies sparse activation to the features using a feedforward neural network. Furthermore, a gated fusion mechanism is incorporated to deeply integrate original intramodal and frequency-domain features, ensuring stable information propagation across deep network layers. Experimental results on three publicly available benchmark datasets demonstrate that AFDSE outperforms existing state-of-the-art methods. This work offers new insights into frequency-domain feature learning for multimodal RS data and highlights the effectiveness of a deep fusion strategy across spatial, spectral, and frequency domains. The code is available at https://github.com/2001weixiao/AFDSE