Fig. 3: Model performance, feature importance and selected features’ relationships with the distance at the threshold of 30 mm/3 hours (DIST30) based on the SHapley Additive exPlanations (SHAP) values from the global XGBoost model.
From: Global expansion of tropical cyclone precipitation footprint
Models are developed based on monthly averaged DIST30 and environmental variables within 0.25° × 0.25° grid boxes. a The scatter density plot of observed and predicted based on 5-fold random cross-validations: one XGBoost model is trained using the data exclusively for each fold and used for prediction based on data from each fold; The predicted values are collected for all five folds and compared with all observed values; color represents the frequency of observations/predictions in each 1/100 bin within the observed DIST30 range. b feature importance calculated as the SHAP value for eight top-ranked features with the most importance for the XGBoost model. c–f Relationship between the four most important features and DIST30 SHAP value, with distributions of each feature. Dot colors in (c–f) denote that the feature has the largest covariance with the main feature (x-axis) in the model.
