Moscow University Soil Studies Bulletin. 2026. 1.

Digital soil mapping (DSM) incorporates the most advanced approaches and methods in the computer analysis and modeling of spatial soil data. Currently, over 90% of all publications in the field of soil mapping are related to DSM. An analysis of scientometric databases and publications shows that the number of publications is growing annually, yet the declared quality of the resulting soil maps remains virtually unchanged. This is due to a number of factors, and the purpose of this article is to identify them through a review of scientific publications. The conducted research has shown that maps created using DSM approaches vary in quality. The accuracy of the models used to create digital soil maps, as declared by the authors, typically ranges from 30–40% to 70–80%. Sixteen potential sources of error in digital soil maps were identified, grouped into errors in the initial data on soil properties, initial data on predictors, errors in model selection, and errors in soils as a modeling object. Some sources of error can be eliminated now or in the near future, but there are errors whose elimination seems impossible at the current stage of scientific and technological development. The greatest potential for significant improvement in the quality of digital soil maps of surface soil horizon properties is present, while the least promising is for maps of soil classification names. The presence of errors in soil maps created using digital soil mapping methods should be disclosed during the creation of any soil map, including masking out areas of the map with the greatest errors.

To mitigate the effects of climate change and preserve and enhance soil fertility, considerable attention is being paid to the development of effective land management methods. Reliable forecasts of soil responses to future climate change and anthropogenic impacts are essential for selecting optimal management strategies. Mathematical modeling is a leading method for analyzing and predicting the spatiotemporal variability of soil organic carbon stocks. It is currently being developed in two main directions. The first uses empirical models obtained using digital soil mapping methods. The second direction represents process-based models of the biogeochemical carbon cycle. Each has strengths and weaknesses. Digital soil mapping models are empirical, based only on the analysis of data on soil properties and environmental characteristics, and therefore are limited in explaining the spatial variability of soil carbon stocks. The uncertainty of forecasts based on these models depends on the volume and quality of the training data. They demonstrate the superiority of process-based models in predicting the spatial distribution of soil organic carbon stocks in cases where large, high-quality data sets were used. The advantage of biogeochemical models is that they are based on accumulated soil science knowledge of the processes that determine the carbon cycle, making them effective in studying the mechanisms of soil carbon stock dynamics. However, forecasts using these models, especially at regional and global scales, are characterized by high uncertainty. Currently, ensemble modeling (the integration of various artificial intelligence algorithms used in digital soil mapping with process-based models) is proposed to reduce forecast uncertainty. Effective development of this strategy requires a thorough understanding of the role of key biogeochemical processes in soil organic carbon stock dynamics to improve the conceptual foundations of the models and increase confidence in the predictions. This article discusses the sources of uncertainty in biogeochemical models and the potential use of minimal models to inform the selection of the required set of key processes and mathematical formalisms for their explicit description in models of different spatiotemporal scales. This can reduce the structural uncertainty of nonlinear biogeochemical models.

This study investigates the potential of the WOFOST crop growth simulation model for yield forecasting and crop condition assessment for several key agricultural crops across large extents of Russia over a fifteen-year analysis period (2005–2020) using soil and climatic datasets. Specific emphasis was placed on adapting information from the 1:2,500,000-scale Soil Map (V.M. Fridland edition) to ensure the model's correct performance. Within the study, model parameters for regional hybrids of barley, sunflower, and maize were obtained for several hundred districts within the Russian agricultural belt, including characteristics of phenology, photosynthesis, respiration, and assimilates partitioning. Parameterization was performed using loss functions based on the Brent, Nelder-Mead, and Differential Evolution algorithms to minimize phenology and yield estimation errors. To enable parameterization and independent accuracy assessment, the initial dataset was divided into a training set and a control set in an 80:20 ratio. The regionally parameterized models provided a good match between the error histograms for the training and control sets, with a near-zero systematic bias, while the mean relative absolute deviation of yield estimates ranged from 20% to 26% depending on the crop. Furthermore, it was demonstrated that the model estimates are not only close to the actual yield values for the studied period but also that the model forecast is capable of capturing multi-year trends in the yield dynamics of the studied crops. Finally, the obtained model crop hybrids were used to assess crop state in terms of the deviation of simulated yield from the yield of an optimal season, with the generation of deviation cartogram maps. The results obtained in this study can be used for decision support system of the agricultural sector, including yield forecasting and state assessment of the studied key crops over large areas using simulation modeling, including forecasts made under various climatic scenarios.