<![CDATA[Haze caused by forest and land fires is a serious problem in South Sumatra Province. One mitigation effort that can be made is to improve the accuracy of rainfall predictions, because rain plays an important role in reducing the potential for fires. This study implements machine learning methods, namely XGBoost and ConvLSTM, to predict spatiotemporal rainfall in areas prone to haze. The results show that ConvLSTM is capable of providing better predictions than the baseline, especially during periods of haze, by considering missing data imputation and masking techniques for disrupted satellite conditions. Increasingly apparent climate change in tropical regions has had a significant impact on rainfall patterns, particularly in South Sumatra, which is one of Indonesia's main agricultural and plantation centers. High rainfall variability can lead to the risk of flooding and drought, as well as disrupting productivity in the education, health, and economic sectors. Therefore, a more accurate rainfall prediction approach is needed to support climate adaptation planning and disaster risk mitigation. This study aims to compare the performance of three approaches to daily rainfall prediction, namely the ConvLSTM-based method, XGBoost, and Persistence, using daily observation data from BMKG for the South Sumatra region for the period 1981–2020. The input variables include average air temperature (Tavg), humidity, sunshine duration, and wind speed, while rainfall is used as the prediction target. The analysis was conducted through a time series approach, statistical distribution, and model performance evaluation using the quantitative metrics Root Mean Square Error (RMSE) and Critical Success Index (CSI). The results show that the ConvLSTM model produced the highest accuracy with an average RMSE of 10 mm/day and a CSI of 0.53, which is better than XGBoost (RMSE 12 mm/day; CSI 0.48) and the persistence method (RMSE 15 mm/day; CSI 0.40). Distribution analysis indicates that light to moderate rainfall occurs more frequently, while extreme rainfall occurs sporadically. The correlation heatmap shows that rainfall has a moderate positive relationship with humidity and a negative relationship with solar radiation, while average temperature and wind play a smaller role. The main contribution of this study is to provide empirical evidence that spatiotemporal deep learning methods such as ConvLSTM are superior in modeling the complexity of tropical rainfall dynamics compared to classical machine learning approaches and simple models. These findings can serve as a basis for the development of early warning systems and interactive climate dashboards at the regional level, while enriching the literature on rainfall prediction in tropical regions using an integrative approach.]]>
