Estimates of annual carbon dioxide fluxes from soil in spruce forests of the «ural-carbon» carbon measurement supersite based on incomplete time series with classical regression approaches and machine learning

Annual carbon dioxide flux from soils of different biomes plays a key role in creating global climate models and in analyzing carbon cycles in terrestrial ecosystems. However, there are significant gaps in such studies at a regional scale. Due to the high complexity of obtaining daily soil respiration values, various modeling methods are used. In this work, based on 2760 soil respiration measurements in spruce forests of the «Ural-Carbon» carbon measurement supersite (Middle Urals) carried out in autumn 2021 and from April to October 2022, annual soil respiration values were estimated using classical regression approaches and machine learning. We also investigated the dependence of the results on the complexity of the model (number of predictors) and the methods used (extrapolation by the random forest model and combined approaches to estimate winter CO₂ fluxes). The «simplified» model with 7 predictors showed only a slight decrease in accuracy compared to the full model with 21 predictors (R² = 0.89,MSE= 0.31 vs.R² = 0.92,MSE= 0.22). Predictors based on remote sensing turned out to be more significant for the accuracy of the model than data measured in the field. Although the initial results of different approaches varied, adding winter respiration values taken from the literature to the random forest model and averaging the values of the combined approaches allowed us to achieve similar values of annual soil respiration: 830.3 ± 6.4 and 851.6 ± 8.0 g C/m²year, respectively.