Abstract: Hyperspectral remote sensing has emerged as a novel technology for monitoring soil environments, the core of which lies in the extraction of characteristic spectral features and the construction of inversion models. However, the cross-regional adaptability of inversion models remains under-studied. In this study, hyperspectral data were pre-processed through spectral transform. The characteristic bands of soil copper were identified using Shapley additive explanations (SHAP). Employing the extreme gradient boosting (XGBoost) and random forest (RF) models, this study conducted hyperspectral inversion and comparative analysis of copper content in soil samples from Fuzhou City. The cross-regional adaptability of the SHAP-XGBoost model was validated in Licheng District, Putian City. The results show that in the modeling area (Fuzhou City), the differential transformation of the raw hyperspectral data could significantly enhance the spectral features. The characteristic bands of soil copper identified using SHAP included 400-500 nm and 1900-2300 nm. The SHAP-XGBoost model proved to be the optimal inversion model for copper content in soils, achieving a coefficient of determination (R²) of 0.73, a root mean square error (RMSE) of 6.4, a mean absolute error (MAE) of 5.38, and a residual prediction deviation (RPD) of 1.92. Compared to the light gradient boosting machine (LightGBM), RF, and partial least squares regression (PLSR) models, the SHAP-XGBoost model improved R² by 4.3%, 17.7%, and 49.0%, respectively. In the validation area (Licheng District), the inversion results of the SHAP-XGBoost model for soil copper showed a R² value of 0.75, a RMSE of 5.01, and a RPD of 2.01. Uncertainty analyses reveal that the average coefficients of variation in Fuzhou City and Licheng District were both below 10%, indicating the strong robustness and excellent cross-regional adaptability of the SHAP-XGBoost model.