TFFNet: Transform Fusion Fuzzy Network for Multimodal Remote Sensing Classification

Multimodal remote sensing classification benefits from the complementary properties of different sensing modalities. However, most existing methods prioritize spatial and channel-wise features, with cross-modal transform-domain correlations often receiving insufficient attention. This imbalance, to some extent, restricts their representational and discriminative capabilities. To address this, we propose a novel fuzzy classification framework that jointly models multiple transform domains to enhance multimodal feature integration and uncertainty handling. First, the discrete Fourier transform (DFT) is applied separately to the image information of each modality, enabling the extraction of global semantic information from different sources. Meanwhile, the proposed wavelet transform (WT)–Mamba module, which integrates selective state-space modeling (SSM) with WT, is designed to capture both long-range dependencies and fine-grained spatial details. To effectively integrate features from the two transform domains, we design a squeeze-and-excitation (SE) transform fusion module that adaptively reweights them based on interchannel correlations, thereby suppressing redundant information and enhancing discriminative features. Finally, the fused features are fed into a fuzzy classifier that integrates Gaussian membership functions and learnable fuzzy rules, enabling robust uncertainty modeling and interpretable decision-making.