Spatiotemporal lagging of predictors improves machine learning estimates of atmosphere–forest CO2 exchange

Abstract

Abstract. Accurate estimates of net ecosystem CO2 exchange (NEE) would improve the understanding of natural carbon sources and sinks and their role in the regulation of global atmospheric carbon. In this work, we use and compare the random forest (RF) and the gradient boosting (GB) machine learning (ML) methods for predicting year-round 6 h NEE over 1996–2018 in a pine-dominated boreal forest in southern Finland and analyze the predictability of NEE. Additionally, aggregation to weekly NEE values was applied to get information about longer term behavior of the method. The meteorological ERA5 reanalysis variables were used as predictors. Spatial and temporal neighborhood (predictor lagging) was used to provide the models more data to learn from, which was found to improve considerably the accuracy of both ML approaches compared to using only the nearest grid cell and time step. Both ML methods can explain temporal variability of NEE in the observational site of this study with meteorological predictors, but the GB method was more accurate. Only minor signs of overfitting could be detected for the GB algorithm when redundant variables were included. The accuracy of the approaches, measured mainly using cross-validated R2 score between the model result and the observed NEE, was high, reaching a best estimate value of 0.92 for GB and 0.88 for RF. In addition to the standard RF approach, we recommend using GB for modeling the CO2 fluxes of the ecosystems due to its potential for better performance.