Attention-driven multi-feature fusion for hyperspectral image classification via multi-criteria optimization and multi-view convolutional neural networks

Hyperspectral image (HSI) classification plays a critical role in various practical applications, including precision agriculture, environmental monitoring, and urban planning, where accurate identification of materials based on their spectral signatures is essential. However, the high dimensionality of hyperspectral data poses significant challenges. HSI classification is a complex problem due to the large number of spectral bands and the limited labeled samples. To address these challenges, we propose an approach that incorporates a spectral feature selection method based on multi-criteria optimization with Three Dimensions (3D) Convolutional Autoencoders (CAE) for spatial feature extraction, followed by a multi-view 3D Convolutional Neural Network (3DCNN) for the spectro-spatial classification by fusing spectral and spatial features. Our spectral feature selection method is designed to identify representative spectral features while minimizing redundancy among bands, thereby retaining essential information for classification. Subsequently, we develop a 3DCAE model for spatial feature extraction, allowing the model to capture spatial information effectively. The 3DCNN model is employed for multi-view fusion, enabling the integration of spectral and spatial information for spectro-spatial classification. By incorporating spectral and spatial features, our approach aims to enhance the classification accuracy of high-dimensional images. We evaluate the effectiveness of our proposed approach on real high-dimensionality image datasets, including Indian Pines, Pavia University, Salinas, and Houston2018. Experimental results demonstrate the superior classification performance of our method, highlighting the efficacy of the multi-criteria optimization model for spectral feature selection and the fusion capabilities of the 3DCNN for spectro-spatial classification.