Runoff simulation based on landscape pattern classification and machine learning

Middle Yellow River Basin, China This study examines how integrated landscape structure affects runoff across multiple scales under climate and human pressures. 37 subbasins were classified using AutoClass, and coupled landscape pattern–runoff models were developed with Partial Least Squares Regression (PLS), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Compared with the runoff model that only considers meteorological factors, the runoff model coupled with landscape patterns has higher accuracy. Among them, the performance of XGBoost is the most outstanding (NSE = 0.966, NRMSE = 0.04；NSE = 0.820, NRMSE = 0.08). SHAP analysis revealed distinct dominant drivers between landscape types, indicating structural–functional differentiation in runoff generation. An LSTM-based classification model achieved 100 % test accuracy and was successfully transferred to ungauged basins, demonstrating strong generalization. This study establishes a transferable classification–coupling framework integrating landscape structure into hydrological modeling, providing new insights for runoff prediction and watershed management. This study proposes a classification-based runoff simulation framework integrating landscape pattern clustering with machine learning. By classifying 37 subbasins in the Middle Yellow River Basin using AutoClass, input heterogeneity was reduced and modeling stability improved. A baseline meteorology-only model and a landscape-meteorology coupled model were compared to evaluate the added value of landscape information. Results show that the XGBoost-based coupled model achieved the best performance (NSE 0.966, NRMSE 0.037), demonstrating strong accuracy and robustness. SHAP analysis revealed key variables and threshold effects across landscape types, and an LSTM-based classifier achieved 100 % accuracy and successful transfer to ungauged basins. This framework offers a robust and transferable modeling pathway for hydrological prediction and watershed management in data-scarce regions. • A classification-based runoff simulation framework was developed integrating landscape patterns and machine learning. • A baseline meteorology-only model confirmed the added value of landscape classification. • The coupled landscape pattern XGBoost runoff model achieved the highest simulation accuracy and the strongest robustness. • SHAP analysis revealed key variables and threshold effects across different basin types. • The framework showed strong spatial transferability and practical watershed management value.