Machine Learning-AIded Groundwater Level and Co2 Emission Estimations in Indonesian Tropical Peatland

Tropical peatlands store a large fraction of global soil carbon and emit large amounts of CO 2 from peat decomposition and root respiration. Accurate quantification of peatland carbon emissions requires a prediction that is sensitive to environmental gradients. This study estimated biweekly total and heterotrophic peat CO 2 emissions (Rs and Rh, respectively) in the peatland hydrological unit (PHU), which represents the whole peatland ecosystem on Rupat Island, Indonesia. The Rs and Rh were estimated using nonlinear CO 2 -to-groundwater level (GWL) response curves derived from 97 observations across 32 Indonesian chamber studies. The GWL dataset was predicted using Extreme Gradient Boosting (XGB), integrating dense GWL measurements (16,496 data points; Jan-2019 to Apr-2025) with dynamic and static predictors. SHAP analysis revealed that surface elevation, distance to the coast, and antecedent climate proxies contribute primarily to the GWL model's predictions. Spatial upscaling of predicted GWL (−35.93±4.67 cm) using Gompertz CO 2 -GWL response curves (P<0.001; R 2 = 0.22-0.24) yielded estimated annual cumulative Rs and Rh of 4.44±0.15 and 3.55±0.11 Mt CO 2 y -1 , totaling 31.08 and 24.83 Mt CO 2 , respectively, over the study period. Pooled results showed interannual and seasonal variations in Rs and Rh of about 3-5 Kt CO 2 biweekly -1 . Cultivated and drained shallow peats in coastal backswamps consistently acted as emission hotspots during dry years. Our estimates represent biogenic emissions only, excluding carbon losses from peat fires and fluvial transport. The Rh estimate was comparable to global and national emission factors (EFs) and other response curves. Our study urged more distributed GWL and site-specific emission monitoring and suggested incorporating these emission datasets into future ecosystem-scale CO 2 modeling.