AMS: A hyperspectral image classification method based on SVM and multi-modal attention network

Hyperspectral (HS) image classification technology is increasingly used for identifying land cover categories. However, spectral aliasing restricts its ability to accurately and completely capture land cover features. To overcome this issue, herein, we introduce a classification method that integrates three modules, namely, a convolutional neural network with an attention mechanism (AMCNN), multi-modal cross-layer adaptive fusion encoder (MCAFE) and support vector machine (SVM), which is referred to as attention-based multi-modal crosslayer fusion network with SVM (AMS). In particular, AMCNN integrates convolution and attention mechanisms to overcome the limitations of a single CNN structure in dynamically allocating attention. MCAFE is proposed to overcome the issues of ineffective inter-layer information interaction and gradient vanishing commonly observed in stacked encoder layers structures. Furthermore, SVM is used to obtain the decision boundaries because of its better performance on linearly separable data than on traditional fully connected (FC) layers. Experimental results demonstrate that AMS considerably enhances the overall accuracy (OA), average accuracy (AA) and Kappa metrics on the Houston and MUUFL datasets, outperforming other state-of-the-art methods.