The research papers assessing the content of soil organic carbon with the help of Vis-NIR spectroscopy approaches are systematically analyzed and subject to meta-analysis. This meta-analysis included 134 studies published in 1986–2022 with a total sample of 709 values of quantitative metrics. The papers have been searched for in databases of scientific periodicals (RSCI, Science Direct, Scopus, and Google Scholar) by the key word combination “Vis-NIR spectroscopy AND soil organic carbon”. The meta-analysis using the nonparametric one-sided Kruskal–Wallis variance analysis in conjunction with nonparametric pairwise method shows the presence of a statistically significant difference between the median values of the accepted quantitative metrics of the predictive power of the models, namely, coefficient of determination ( R ² _cv/val ), root mean square error (RMSE), and the ratio of performance to deviation (RPD). The best performance of the preprocessing method for spectral curves is demonstrated and the estimates of soil organic carbon content obtained by laboratory and field spectroscopies are compared.

The implementation of the sustainable management of the interaction between agriculture and the environment requires an increasingly deep understanding and numerical description of the soil genesis and properties of soils. One of the areas of application of relevant knowledge is digital irrigated agriculture. During the development of such technologies, the traditional methods of soil research can be quite expensive and time consuming. Proximal soil sensing in combination with predictive soil mapping can significantly reduce the complexity of the work. In this study, we used topographic variables and data from the Electromagnetic Induction Meter (EM38-mk) in combination with soil surface hydrological variables to produce cartographic models of the gravimetric soil moisture for a number of depth intervals. For this purpose, in dry steppe zone conditions, a test site was organized. It was located at the border of the parcel containing the irrigated soybean crop, where 50 soil samples were taken at different points alongside electrical conductivity data (EC_a) measured in situ in the field. The modeling of the gravimetric soil moisture was carried out with the stepwise inclusion of independent variables, using methods of ensemble machine learning and spatial cross-validation. The obtained cartographic models showed satisfactory results with the best performance R²_cv 0.59–0.64. The best combination of predictors that provided the best results of the model characteristics for predicting gravimetric soil moisture were geographical variables (buffer zone distances) in combination with the initial variables converted into the principal components. The cartographic models of the gravimetric soil moisture variability obtained this way can be used to solve the problems of managed irrigated agriculture, applying fertilizers at variable rates, thereby optimizing the use of resources by crop producers, which can ultimately contribute to the sustainable management of natural resources.

This paper discusses the effectiveness of using a number of methods of preprocessing spectrometric data in the 325–1075-nm wavelength range in order to predict the concentration of organic soil carbon. Methods of preprocessing spectral data (moving-average filtering, Savitzky–Golay smoothing, calculating the first and second derivatives, and scaling) were consecutively applied to the spectral data of soils (with their natural texture and pulverized) to increase the reliability and effectiveness of the models. According to the criterion of maximizing the determination coefficient and minimizing the rms error as a cross check, the best method of predicting organic soil carbon turned out to be partial-least-squares regression with computation of the first derivatives of the original spectra (Rcv2=0.758RMSE_cv=0.492).

Приведены результаты цифрового картографирования содержания органического углерода в пахотных горизонтах почв и оценки точности получаемых моделей с использованием методов машинного обучения для участка Среднерусской возвышенности Воронежской области. Цифровое картографирование основывалось на 22 точках почвенного опробования, используемых для обучения и проверки моделей, а также на нескольких наборах переменных-предикторов, в качестве которых выступали цифровая модель рельефа, производные от нее и данные дистанционного зондирования различного пространственного разрешения. Для построения моделей пространственного варьирования исследуемого свойства использовали несколько методов, основанных на деревьях решений: ансамбль деревьев решений, бустинг регрессионных деревьев и байесовские регрессионные деревья. Оценку точности полученных картографических моделей определяли методом перекрестной проверки, при этом в качестве показателей точности использовали коэффициент детерминации, среднюю абсолютную ошибку и корень среднеквадратичной ошибки. По результатам моделирования выявлено, что с использованием переменных-предикторов, представленных цифровой моделью рельефа, ее производными и данными Landsat 8 удалось получить более устойчивые модели, причем коэффициент детерминации изменяется от 0.6 до 0.7, RMSE_cv, т.е. ошибка прогноза от 0.5791 до 0.6520. Лучшая модель получена с помощью метода байесовских регрессионных деревьев; тогда как для переменных-предикторов, представленных цифровой моделью рельефа, ее производными и данными Sentinel 2 - от 0.47 до 0.55, ошибка прогноза от 0.7031 до 0.7909. Выявлено, что в описанных моделях по различным наборам данных наиболее значимыми оказывались разные переменные-предикторы.

This study presents an approach for predicting soil class probabilities by integrating synthetic composite imagery of bare soil with long-term vegetation remote sensing data and soil survey data. The goal is to develop detailed soil maps for the agro-innovation center “Orlovka-AIC” (Samara Region), with a focus on lithological heterogeneity. Satellite data were sourced from a cloud-filtered collection of Landsat 4–5 and 7 images (April–May, 1988–2010) and Landsat 8–9 images (June–August, 2012–2023). Bare soil surfaces were identified using threshold values for NDVI (<0.06), NBR2 (0.10). Synthetic bare soil images were generated by calculating the median reflectance values across available spectral bands. Following the adoption of no-till technology in 2012, long-term average NDVI values were additionally calculated to assess the condition of agricultural lands. Seventy-one soil sampling points within “Orlovka-AIC” were classified using both the Russian and WRB soil classification systems. Logistic regression was applied for pixel-based soil class prediction. The model achieved an overall accuracy of 0.85 and a Cohen’s Kappa coefficient of 0.67, demonstrating its reliability in distinguishing the two main soil classes: agrochernozems and agrozems. The resulting soil map provides a robust foundation for sustainable land management practices, including erosion prevention and land use optimization.