Mixed convolutional classification method for hyperspectral images based on spatial spectrum orthogonal constraints and bidirectional attention mechanism

When classifying hyperspectral images, 3D-CNN and 2D-CNN are limited in their application due to excessive consumption of computing resources and difficulty in effectively extracting image features. Although the emergence of mixed convolution can reduce some limitations to a certain extent, there are also some issues. (1) After performing mixed convolution on images, there may be phenomena of information disappearance or redundancy. (2) In the case of small samples, classification accuracy will significantly decrease. (3) Maximum pooling is usually used in feature compression processing, but it ignores the relationships between features. To address these issues, this paper proposes a parallel convolution architecture and introduces spatial spectral orthogonal constraints and bidirectional attention mechanism to ensure the independence of spectral spatial information in mixed convolution and enhance the utilization of key information to solve problems 1 and 2, respectively. In terms of feature compression processing, this method considers combining max pooling and covariance pooling to balance salient features and high-order statistical features, enhancing the representation ability of features. For data output, this method will use the Mish function with better stability to replace the commonly used ReLU function. Experiments on three publicly available datasets showed that the OA values of the proposed method were 93.3%, 94.1%, and 99.1%, respectively. It is significantly superior to commonly used classification methods. In short, the proposed method demonstrates powerful discriminative features and improves the classification performance of images.