<p>Machine learning (ML) models are increasingly used for Tropical Cyclone (TC) intensity forecasting. However, their black-box nature limits operational adoption. Explainable AI (XAI) methods are essential for building interpretability, yet standard approaches such as SHAP often disrupt the spatial coherence fundamental to geophysical data. GeoShapley mitigates this by treating geographic location (GEO) as a single player, but its sampling-based estimator is computationally expensive and, as shown here, fails to isolate the main spatial effect. This study introduces Geo-dSHAP, a computationally efficient, gradient-based attribution framework that separates the main geographic contribution from feature–interaction effects. We apply Geo-dSHAP to six ensemble models. XGBoost, LightGBM, Random Forest, HistGradientBoosting, Gradient Boosting, and ExtraTrees, trained to predict 24-hour TC intensity change in the Western North Pacific. Geo-dSHAP quantifies a substantial main effect for the GEO player (5–8% importance). Moreover, the spatiotemporal analysis of the best-performing model (XGBoost) shows that it has learned complex, non-linear physical relationships. Feature sensitivity is non-monotonic–thermodynamic potential (POT) peaks during the Decay phase as the storm recurves, and the importance of leading predictors reaches a maximum for the strong TY (Typhoon) category before declining for the most intense super typhoons. Geo-dSHAP thus provides a diagnostic framework for model evaluation beyond accuracy metrics. By identifying where and when models express physical sensitivity, it delivers physically interpretable insights that improve trust and transparency in ML-based operational forecasting.</p>