Temporally Robust Rainfall Classification from Radiosonde-Derived Atmospheric Instability Indices: A Machine Learning Approach with SHAP Interpretability

We evaluate binary rainfall occurrence classification in southern Thailand using 00 UTC radiosonde profiles from Phuket Airport and rainfall observations at Phuket, Krabi, Takua Pa (Phang-nga) and Nakhon Si Thammarat. Of 2631 soundings from 2011 to 2021, 1487 met quality-control and layer-continuity criteria. Five classifiers (logistic regression, support vector machines, random forests, extreme gradient boosting, and multilayer perceptrons) were evaluated using year-grouped cross-validation (2011–2019) and a chronological holdout (2020–2021). Holdout receiver operating characteristic area under the curve (ROC-AUC) reached 0.845 at Phuket, 0.793 at Krabi and 0.873 at Takua Pa, exceeding station-specific climatological baselines. At Nakhon Si Thammarat, the maximum holdout ROC-AUC was 0.691, indicating reduced spatial representativeness. Tree-based Shapley Additive Explanations (TreeSHAP) identified the K-index as the leading thermodynamic contributor across western stations, with geographically varying secondary contributions. Because daily rainfall windows include pre- and post-sounding periods, separate six-hour post-sounding evaluations (07:00–13:00, 13:00–19:00 and 19:00–01:00 local time) were conducted. Evening discrimination was consistently lower than morning discrimination across station–classifier combinations. A localized terrain–wind proxy provided negligible improvement. These results support temporally validated, station-specific rainfall occurrence classification while highlighting spatial representativeness and lead-time limitations for prospective operational use.