Abstract: Using the Huangfengqiao Forest Farm in Hunan Province as the study area, based on Zhuhai-1 and Sentinel-2 images from summer and autumn, an improved differential evolution algorithm integrating social learning and environmental pressure mechanisms was used for band selection; spectral, principal component, and texture features were extracted and fused; the SHAP–RFE method was used for optimization; and three algorithms, XGB, GBM, and RF, were used for dominant tree species classification. The results showed that among six optimization algorithms for band selection on Zhuhai-1 data, SLEPDE achieved the best performance, with band reduction rates of 78.13% in summer and 68.75% in autumn. Through the SHAP–RFE method, the classification accuracy was further improved. Using Zhuhai-1 data, the highest overall classification accuracy reached 91.40% in summer and 91.07% in autumn, outperforming the results obtained from Sentinel-2 data during the same period. SHAP analysis showed that bands such as B25 and B19 contributed most significantly to model performance, and the importance of key features varied among different tree species. Zhuhai-1 hyperspectral imagery has great potential in dominant tree species classification. The SLEPDE algorithm reduces hyperspectral data redundancy and achieves high classification accuracy with fewer bands. The SHAP–RFE method improves model performance, reveals the features driving classification of different dominant tree species, and enhances the transparency and credibility of model decisions.