Performance Analysis of CatBoost, XGBoost and Random Forest Models for Cloud Top Height Predicition Using Satelliite Data

Abstract

Cloud Top Height (CTH) is a critical parameter for convective weather monitoring and aviation meteorology, particularly over the Indonesian maritime continent, where deep convection occurs frequently. This study evaluates the predictive performance and behavior of three tree-based ensemble models XGBoost, CatBoost, and Random Forest—for estimating CTH from Himawari multispectral observations using the Sentinel-5P/TROPOMI CTH product as the reference. The input variables comprise Himawari brightness-temperature channels, cloud fraction, latitude, and longitude. Evaluation on an independent spatial validation set of 3,444,543 observations shows that XGBoost achieved the best aggregate performance, with a mean absolute error of 971.71 m, root mean square error of 1297.57 m, coefficient of determination of 0.646, and mean bias of -33.55 m. CatBoost and Random Forest produced RMSE values of 1340.95 m and 1371.11 m, respectively. Multi-method interpretability analysis using built-in importance, permutation importance, group permutation, SHAP, partial dependence, and individual conditional expectation indicates that spectral channels and cloud fraction are the principal information sources, whereas coordinates provide weaker but non-zero spatial context. The error audit reveals relatively stable predictions for mid-level clouds but strong underprediction for extremely high cloud tops. Inter-model disagreement is positively but weakly related to actual error, indicating that it is a useful diagnostic signal rather than a calibrated uncertainty estimate. These results demonstrate that model selection for operational meteorology should consider not only aggregate accuracy but also error regimes, spatial dependence, and physical consistency.