2026 , Volume 31, № 5, p.54-76

The purpose of this work is to present new results of calculations of gross primary production (GPP) for the territory of Southern Western Siberia, obtained from the MODIS spectroradiometer data, ERA5 reanalysis and the SRTM digital elevation model using machine learning methods. GPP is an important indicator of carbon exchange between the atmosphere and the ground surface, and quantitatively characterizes the growth of plant biomass during the vegetation season.

Methods. Six machine learning methods are used to obtain GPP estimates: random forest, AdaBoost, XGBoost, LightGBM, gradient-boosted decision trees, and a neural network model based on a multilayer perceptron. The training and test sets of the models for the 2024 vegetation season include 14 features to which, according to independent studies, gross primary production is the most sensitive. The target dataset for machine learning was formed using data from the MOD17A2H product, which contains GPP estimates with a resolution of 500 m and a time step of 8 days. The final dataset includes data for the May–September 2024 vegetation season.

Results. The most optimal model configurations were established. An analysis of the model training quality, which carried out at nine locations across all natural and climatic zones of the study area, has showed that all models adequately reproduce the seasonal variation of GPP during the 2024 vegetation season. As a result of the analysis of regression metrics, it was found that the LightGBM model of aggregate learning based on random forest with gradient boosting has the best quality of GPP predictions. An analysis was conducted to evaluate the impact of features that most significantly influence the quality of GPP modelling using the LightGBM model. The results revealed that the features most frequently acting as decision determinants for the model — specifically land surface temperature (LST), vegetation indices (EVI and NDVI), and the land surface water index (LSWI) — make a relatively minor contribution to minimizing the error function (i. e., improving prediction quality). The primary contribution to reducing the error function comes from the fraction of absorbed photosynthetically active radiation (FPAR) and the leaf area index (LAI); together, these two features served as the determining factors in the construction of decision tree nodes in more than 60 % of cases.

Conclusions. Machine learning methods can be effectively used to determine quantitative GPP estimates for the southern part of Western Siberia. However, more sophisticated methods based on neural networks or more complex aggregation strategies are less effective for this task.