OCSA-FN: A Fusion Network With Orthogonality-Constrained Spatial Attention for Hyperspectral and Land Surface Temperature Data Classification

With the continuous advancement of remote sensing (RS) technology, collaborative image land cover classification using multi-source RS data has gradually emerged as a prominent research focus. However, one of the core challenges during the feature fusion process of multi-source RS data is to achieve efficient feature interaction while minimizing redundancy and enhancing the representational capacity of features. To address this issue, this paper proposes a fusion network with orthogonality-constrained spatial attention (OCSA-FN) for hyperspectral image (HSI) and land surface temperature (LST) data classification. First, OCSA-FN employs a dual-branch convolutional neural network (DB-CNN) module, where one branch utilizes a cross spatial-spectral CNN (CSS-CNN) to extract spectral features from HSI, while the other branch incorporates a learnable Sobel CNN (LS-CNN) to adaptively extract temperature features from single-channel LST data. Next, OCSA-FN introduces a dual-pooling residual channel attention (DRCA) module that leverages pooling-based interaction and residual connect to perform channel-wise weighting on deep features. Subsequently, OCSA-FN presents an adaptive orthogonal feature fusion (AOFF) module designed to construct two sets of mutually orthogonal spatial basis vectors by imposing orthogonal constraints, this effectively reduces the feature redundancy. Meanwhile, an adaptive spatial attention mechanism dynamically adjusts the fusion weights of the features between HSI and LST, facilitating efficient complementary fusion of multi-source features. Finally, the weighted fused features are utilized for the classification. Extensive experimental results demonstrate that OCSA-FN outperforms state-of-the-art existing methods in terms of classification accuracy while reducing both the parameter count and computational complexity within the model.