Eddy Covariance-Informed Estimation of Evapotranspiration using Machine Learning in Tropical Peatland Cultivated by Oil Palm Plantation

Although numerous studies have documented machine learning/ML applications in estimating evapotranspiration/ET at the ecosystem scale, research on this theme in tropical peatland is considered minuscule, particularly using continuous, long-term data of eddy covariance/EC at oil palm plantations/OPP. This research presents an initial exploration of several ML algorithms in estimating ET fluxes using 11,694 half-hourly data of micrometeorological and soil measurements collected from an EC tower installed in OPP-cultivated peatland in South Kalimantan, Indonesia. We evaluated five ML (Cubist, tree regression/TR, random forest/RF, gradient boosting machine/GBM, extreme gradient boosting/XGB, and a shallow artificial neural network/ANN) and three GLM (multiple linear regression/MLR, regularized-GLM/GLMnet, and Poisson-GLM/GLMpois) algorithms performances that were trained from 11 environmental predictors. Cubist and RF obtained the best performances (Rsq = 0.90; MAE = 0.04; RMSE = 0.07) and outperformed GLM-based models by up to 73%. Generally, the entire ML models achieved the best predictive performances, but slightly above GLM models, except for TR. Photosynthetic-driven radiation was the most influential contributor to estimating ET, whereas moderate to low importances were obtained by wind, heat, and soil components. This study supports previous findings emphasizing the importance of data-driven ML modelling in predicting high frequency-informed ET from the EC system in tropical peatlands cultivated by OPP.