<![CDATA[Rainfall prediction plays an important role in supporting hydrometeorological disaster mitigation and weather-related decision-making. However, accurate rainfall prediction remains challenging because atmospheric processes are highly nonlinear and governed by complex interactions among multiple meteorological variables. This study proposes a Hybrid TabNet–XGBoost model for daily rainfall prediction using integrated radiosonde and surface meteorological observations collected at the BMKG Juanda Class I Meteorological Station. The dataset covers the period from 2019 to 2025 and consists of 2,551 daily observations. TabNet was employed to select the fifteen most informative atmospheric variables based on feature importance, while temporal dependencies were incorporated through lag features generated using Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) analyses. Hyperparameter optimization was performed using Optuna with TimeSeriesSplit cross-validation prior to model training. Experimental results on the testing dataset achieved an RMSE of 19.1347 mm, an MAE of 11.9742 mm, a MAPE of 17.62%, and an R² of 0.0449. The proposed model was able to capture the general temporal pattern of daily rainfall and produced satisfactory predictions under the dominant rainfall conditions represented in the dataset. However, the model exhibited reduced sensitivity to high-intensity rainfall events, resulting in the underestimation of extreme rainfall and a relatively low R² value, primarily due to the imbalanced rainfall distribution and the complexity of rainfall processes. The optimized model was subsequently applied to generate daily rainfall projections for 2026 based on historical atmospheric observations. Since the corresponding observational data were unavailable at the time of this study, these projections should be interpreted as model-based forecasts rather than validated prediction results. Overall, the proposed Hybrid TabNet–XGBoost framework demonstrates the potential of integrating radiosonde and surface meteorological observations for daily rainfall prediction while highlighting the need for additional atmospheric and spatial information to improve the prediction of extreme rainfall events.]]>
