Abstract: The rapid development of hyperspectral remote sensing technology in China fully ensures the effective application of large-scale surface feature classification. However, achieving high-precision classification under few-spot conditions while fully leveraging hyperspectral spatial-spectral information remains challenging. This study developed a 3D convolutional autoencoder (3D-CAE) network guided by physical constraints from mixed pixel decomposition. This approach enables accurate estimation of endmember abundance while effectively expressing regularized spatial-spectral features of hyperspectral data. In combination with a support vector machine (SVM) classifier, the method achieves hyperspectral classification under few-spot conditions. The classification performance of various models was evaluated at different sampling rates. To validate the proposed method, this study conducted experiments including comparisons with traditional hyperspectral feature extraction and classification methods, such as supervised classification approaches. The classification performance of various models was also evaluated at different sampling rates. The experimental results demonstrate that the proposed hyperspectral classification method has a significant advantage of accuracy, achieving a mean intersection over union (mIoU) of 0.829, which was close to 0.8 even at a low sampling rate of 1/200, surpassing its counterparts. These results confirm that the proposed method exhibits robustness under few-spot conditions. This study provides a valuable technical reference for addressing hyperspectral classification challenges under few-spot conditions.