This study is based on ERA5 reanalysis data and develops an interpretable machine learning framework combining Extreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) to investigate the spatial reconstruction and driving mechanisms of 30–90-day intraseasonal precipitation variability over the South Asian monsoon region (5°–25°N, 70°–95°E). The model incorporates 12 tropical monsoon-related predictors, representing atmospheric dynamics, vertical motion, moisture conditions, and air–sea coupling processes. Different lag-time windows (lag5–lag20) and neighborhood spatial averaging schemes are introduced to assess the time-scale dependence of precipitation predictability.

The results show that the model effectively reconstructs the spatial patterns of intraseasonal precipitation anomalies and achieves strong predictive performance on independent test data. Model skill decreases with increasing lead time, whereas incorporating spatial neighborhood information significantly reduces prediction errors, indicating that intraseasonal precipitation variability depends on both temporal memory and local dynamical–moisture structures.

SHAP analysis reveals that 500 hPa vertical velocity (w500) is the dominant controlling factor across all lag scales, highlighting the key dynamical role of mid-level vertical motion in monsoon oscillations. Moisture conditions in the 700–850 hPa layer and sea surface temperature (SST) also exert important influences, while horizontal wind and convergence–divergence mainly affect precipitation indirectly through coupling with vertical motion. With increasing lag time, the contribution of moisture-related factors increases, while that of dynamical factors, particularly vertical velocity, weakens, suggesting a transition from a dynamics-dominated regime to a thermodynamic–moisture-controlled regime.

Spatial SHAP analysis further indicates clear regional contrasts between the Indian Peninsula and the Bay of Bengal. The Bay of Bengal is more sensitive to vertical motion and air–sea coupling, whereas the Indian Peninsula is more strongly influenced by moisture transport and low-level winds. SHAP-based dependency analysis also reveals a nonlinear interaction between vertical velocity and moisture, showing that strong ascent can significantly amplify precipitation anomalies under high-humidity conditions.

Overall, this study demonstrates that the XGBoost–SHAP framework can effectively characterize both the spatial structure and physical drivers of intraseasonal precipitation variability in the monsoon region. It improves predictive skill while providing physically interpretable insights into dynamical–moisture coupling, offering new machine learning-based evidence for understanding multiscale modulation of the South Asian monsoon.