Hierarchical Receptive-Field Selection with Attention ResNet for Hyperspectral Image Classification

Hyperspectral Images (HSIs) have a high spectral resolution, employed in various applications, including materials discrimination in remote sensing. However, spectral bands suffer from noise and spectral correlation, limiting the typical selection of spatial-spectral kernels. This is addressed using Three Dimensional Convolutional Neural Networks (3D-CNNs) with fixed Receptive Fields (RFs). However, these approaches cannot effectively select the appropriate RF during network design. To address this, a Hierarchical Receptive-Field Selection with Attention ResNet (HRSAResNet) is proposed for HSI classification. The proposed network is composed of a proposed 3D Function-Mixture (FM), consisting of basis functions, i.e. multi-branched sub-nets with different RFs and an additional sub-net as a mixing function. This approach adaptively determines the appropriate RF branch and the mapping function at the pixel level. An enhanced 3D Residual Block (ResBlock) enriched with a Channel Attention Module (CAM) is proposed to exploit inter-channel feature dependencies, boosting the classification performance. Extensive experiments are performed on three datasets; IP, KSC, and UP. The comparative results demonstrate the superior performance of the proposed method by achieving a maximum of relative improvements up to (2.31%), (4.48%), and (1.66%) in Overall Accuracy (OA), Average Accuracy (AA), and measure the consistency between the obtained and the ground truth maps by Kappa (κ) to the top compared deep learning model.