Monitoring degraded tropical peatland subsidence by integrating SBAS InSAR and machine learning

Abstract In the 1990s, the ambitious yet disastrous 1 million ha Mega Rice Project (MRP) in Central Kalimantan, Indonesia, left behind vast areas of degraded and abandoned peatlands. The drainage of these peat soils has led to subsidence and substantial carbon emissions. Interferometric Synthetic Aperture Radar (InSAR) has been used to quantify peat subsidence in this vast area through satellite data, although it often suffers from incomplete spatial coverage due to decorrelation. In this study, we employed time-series Small BAseline Subset (SBAS) InSAR combined with data-driven machine learning models to estimate peat subsidence in Blocks B and C of the ex-MRP area. A stack of Sentinel-1 C-band data from 2021–2022 served as the primary dataset for SBAS InSAR. To extrapolate InSAR results, we applied and examined several machine learning algorithms by involving some predictor maps. Random Forest Regression (RFR) delivered the best performance when the training data were separated by peatland blocks. The final subsidence map showed mean rates of −1.72 ± 1.57 cm yr –1 (Block B) and −1.55 ± 2.27 cm yr –1 (Block C). Feature importance analysis highlighted peat depth, latest fire, and distance to peat edge as key predictors. Overall, this work demonstrates the potential of integrating InSAR and machine learning to monitor tropical peatland subsidence at landscape scale of peat hydrological units.