<p>Controlling peatland fires is challenging because they can spread throughout the surface and sub-surface, often undetected. This study presents a new perspective through a data-driven approach to estimate potential peatland fires in the form of hotspots and carbon emissions as indicators driven by meteorological and hydrological data. This approach offers deeper quantitative analysis and more practical applications for fire mitigation. Three machine learning-based models, namely neural network (NN), random forest (RF), and light gradient boosting machine (LightGBM), are applied for two stations, Jambu and Pinang Habang, in South Kalimantan, Indonesia. Based on the mean absolute error (MAE), normalized root mean squared error (NRMSE), and correlation coefficient (CC), the best hotspot estimation at Jambu station is shown by the LightGBM model with hydrological data as a predictor. On the other hand, the RF model with hydrological and meteorological data as predictors shows the best performance for emission prediction. At Pinang Habang station, the RF model with hydrological data as a predictor shows the best performance in estimating both hotspots and emissions. Generally, the RF model with only hydrological data as a predictor performs best compared to other models. This result indicates that peatland fires highlight the dominant role of hydrological factors in peatland fire activity, i.e. groundwater levels, soil moisture, and soil temperature.</p>