Abstract: Hyperspectral remote sensing image classification has attracted widespread attention，yet the performance of classification methods remains greatly limited by challenges such as spectral variability （same object with different spectra），spectral confusion （different objects with similar spectra），and limited availability of training samples. To fully exploit the spatial-spectral features of hyperspectral images，this study proposed an improved network integrating residual convolution and neighborhood attention mechanisms. The proposed method consists of：（1） a residual-based spectral feature extraction module combining residual connections and a 3D convolutional neural network （3D-CNN）；（2） a spatial-spectral feature fusion module using mixed convolutions；and （3） a neighborhood attention module designed to enhance the model's ability to focus on homogeneous regions. Experiments were conducted on three public hyperspectral datasets-Indian Pines，Pavia University，and Houston 2013. The results demonstrate that the proposed method achieves higher classification accuracy compared to recent state-of-the-art approaches. Using less than 10% of the samples for training，it attains overall accuracies of 99.39%，99.67%，and 98.64%，respectively，confirming its capability for high-accuracy classification under small-sample conditions.