Aug 2026· Agronomy· Vol 16, pp. 1562· 0 citations· 53 references
Abstract
Soil organic carbon (SOC) is a key indicator of cropland quality, soil fertility, and the carbon sequestration potential of agroecosystems. Accurate characterization of its spatial distribution is essential for black soil conservation and regional soil carbon management. However, regional-scale SOC prediction is often constrained by limited field observations, which can reduce model generalizability and predictive reliability. In this study, we developed a limited-sample SOC prediction framework for the Liaohe Plain using 310 surface (0–20 cm) soil samples collected in 2025 and multi-source environmental covariates, including climate, vegetation, soil spectral, and terrain variables. The framework used the Tabular Prior-Data Fitted Network (TabPFN), whose performance was compared with that of Random Forest, Support Vector Machine, CatBoost, K-Nearest Neighbors, and XGBoost. Model performance was evaluated using 100 repetitions of random 80:20 holdout validation and repeated five-fold spatial cross-validation based on spatially constrained clustering, while sampling-induced relative uncertainty was quantified using 100 repeated random sampling and model-fitting runs. Under random holdout validation, TabPFN showed competitive predictive performance, with mean R2 and RMSE values of 0.608 ± 0.038 and 4.026 ± 0.184 g kg−1, respectively. Repeated spatial cross-validation yielded more conservative performance estimates, with mean R2 and RMSE values of 0.535 ± 0.072 and 4.33 ± 0.31 g kg−1, respectively, indicating that random splitting may overestimate model performance when sampling sites are spatially clustered. Spatial prediction showed that cropland SOC ranged from 4.63 to 27.04 g kg−1, with generally lower values in the west and higher values in the northeast. Areas with high sampling-induced relative uncertainty were mainly concentrated in the northern, northeastern, and marginal regions. These findings provide a methodological basis for SOC mapping, supplementary sampling optimization, and regional soil carbon management under limited-sample conditions, although the temporal robustness of the results requires confirmation using independent data from additional years.
This study evaluates an innovative field-scale targeted sampling strategy within a regional hybrid spatial prediction model that combines machine learning and geostatistics. The framework is designed so that newly collected field observations are incorporated only through the local residual kriging step, while the regional Random Forest trend model remains unchanged, allowing field-scale predictions to be refined without full model refitting. The proposed sampling approach integrates a Normalized Difference Vegetation Index (NDVI)-based Productivity Index, a field-specific index derived from long-term satellite-based NDVI time series, with regional model-derived prediction uncertainty. The approach was evaluated on 28 test agricultural fields containing 7 to 19 in-field soil organic carbon (SOC) observations, which allowed assessment across diverse sampling configurations and levels of within-field variability. The results showed that the regional model alone provided a reasonable baseline prediction (mean root mean square error (RMSE) = 0.24% SOC) but was unable to adequately represent local spatial heterogeneity. Incorporating all available field observations produced the highest accuracy (mean RMSE = 0.17% SOC), while the proposed targeted sampling strategy achieved comparable performance (mean RMSE = 0.18% SOC) using only four strategically selected samples. In contrast, the results also showed that seemingly representative random sampling can lead to poor predictions when informative locations are missed, in some cases performing even worse than the regional model alone. In addition to conventional accuracy metrics, minimum detectable change (MDC) was used to evaluate the capability of the framework to detect meaningful SOC changes beyond prediction error, relevant for SOC monitoring and carbon farming. The results demonstrate that integrating regional models with productivity-based targeted sampling can substantially improve field-scale SOC prediction while avoiding the risks of misleading assessments associated with unfavorable sampling configurations.
Milutin Pejović, M. Kilibarda, D. Protić et al.· Environmental Monitoring & A...· 0 citations
Accurate prediction of soil organic carbon (SOC) at fine resolution is crucial for precision soil management; however, existing national products for Hungary remain too coarse for farm-scale applications. Focusing on Hungarian croplands, we developed a 30 m resolution SOC map using multi-temporal bare-soil composites, DEM derivatives, SHAP interpretability and bootstrap uncertainty. Among five evaluated algorithms, the GBDT model achieved the best performance (test R2 = 0.518, RMSE = 4.498 g·kg−1, MAE = 3.499 g·kg−1, RPIQ = 2.229, LCCC = 0.621). SHAP analysis revealed pronounced nonlinear effects of spectral and topographic variables within this modeling framework, with spectral predictors playing a dominant role in SOC prediction. Furthermore, the bootstrap uncertainty framework yielded a Prediction Interval Coverage Probability of 94.59% at the 95% confidence level, indicating reliable interval estimation for the test set. Spatial patterns of uncertainty varied considerably, with higher values in the western hills and southern sands, and moderate levels in the northern low-mountain areas. Benchmark comparisons showed that our 30 m map captures fine-scale heterogeneity often smoothed over by coarser products, while the uncertainty layer supports risk-aware interpretation. Overall, this study provides a regionally calibrated framework for mapping in similar heterogeneous agricultural landscapes, providing practical insights for local management.
Reliable spatial information on soil organic carbon (SOC) is important for soil-fertility management, climate-resilient coffee production, and land-restoration planning, yet the heterogeneity of smallholder agricultural landscapes is difficult to represent with field observations alone. This study characterised SOC variability, compared four machine-learning (ML) algorithms—Random Forest (RF), Cubist, Gradient Boosted Machines (GBM), and Support Vector Machine (SVM)—identified the main predictive environmental covariates, and generated SOC prediction maps for a smallholder Robusta coffee landscape in South–Central Uganda. A total of 126 georeferenced surface-soil observations (0–30 cm) were analysed for SOC using the Walkley–Black dichromate wet-oxidation method and related to climatic, terrain, spectral, soil, land-cover, and spatial-position covariates within the SCORPAN framework. Models were tuned within an SOC-stratified training subset and evaluated using a withheld 20% random holdout. Observed SOC ranged from 0.31% to 5.30%, with a mean of 2.06% and a coefficient of variation of 62.7%. Independent holdout performance was limited. SVM performed best (R2 = 0.36; RMSE = 1.05), followed by Cubist, RF, and GBM. Annual mean temperature and spatial position had the highest predictive importance, followed by soil type, Landsat 9 Band 5, and annual precipitation. All four models reproduced a broad pattern of relatively higher SOC in Kalungu and central Masaka and lower SOC in Southern Kyotera and parts of Lwengo, although the magnitude of local variation differed among algorithms. The resulting maps are, therefore, best suited for landscape-level screening, sampling prioritisation, and identifying areas requiring field verification, rather than farm-level SOC prescription.
Isaac T. Okurut, Catherine Mulinde, S. D. Ddumba et al.· Land· 0 citations
Soil organic carbon (SOC) stocks are strongly influenced by land use, vegetation condition and climate, particularly in heterogeneous mountain catchments. This study quantified SOC stocks and simulated long-term SOC dynamics in the Çapakçur micro-catchment, eastern Türkiye, by integrating field assessment, geostatistical prediction, uncertainty analysis, inverse RothC calibration and scenario modelling. A total of 428 soil samples were collected from the 0–30 cm layer across forest, degraded forest, and pasture areas. SOC stocks were calculated from SOC concentration, bulk density and soil depth, and spatially predicted using ordinary kriging of log-transformed SOC stocks. RothC was calibrated for each land-use class to estimate the annual carbon inputs required to maintain observed SOC stocks, followed by 50-year restoration and climate-sensitivity simulations. SOC stocks ranged from 7.69 to 247.68 Mg C ha−1, averaging 55.52 Mg C ha−1. Forest had the highest mean SOC stock (78.5 Mg C ha−1), followed by pasture (55.9) and degraded forest (50.2 Mg C ha−1). Required annual carbon inputs were 5.17, 4.58 and 3.29 Mg C ha−1 yr−1, respectively. Increasing degraded forest carbon inputs to forest-equivalent levels increased SOC by 14.76 Mg C ha−1 over 50 years, equivalent to 37.77 Gg C or 138.49 Gg CO2eq at the catchment scale. A stronger restoration scenario increased this potential to 63.77 Gg C. Warming caused SOC losses, with +2 °C reducing catchment SOC by 56.78 Gg C. These findings demonstrate the potential of degraded forest restoration for SOC sequestration while highlighting the vulnerability of long-term SOC gains to climate warming.
Yasin Demir, A. Meral, Azize Doğan Demir· Land· 0 citations
Monitoring soil organic carbon (SOC) stocks is critical for understanding changes in soil health, supporting land management, and informing climate change mitigation policies. This study presents a pragmatic and scientifically robust approach for national-scale resampling of SOC monitoring sites in Australia. Using legacy data from the Soil Carbon Research Program (SCaRP), which surveyed over 4000 agricultural sites between 2009 and 2012, we developed a hybrid resampling framework that combines model-based and design-based sampling principles. A subset of 308 sites comprising 207 ‘standard’ (composited) and 101 ‘detailed’ (replicated) sites (sampling support is 25 m × 25 m) was selected using a two-stage procedure involving k-means clustering, fuzzy allocation, and conditioned latin hypercube sampling. Covariates representing SOC drivers, including climatic, biophysical, and management-related factors, were processed using discounted temporal weights to capture antecedent conditions. The final set of 308 resampled sites (7.57% of SCaRP) reflects the diversity of Australian agroecological conditions and maintains statistical representativeness across key SOC drivers. This was demonstrated through Jensen–Shannon divergence analysis, which yielded values close to zero, indicating high efficiency in the resampling campaign. The resampling approach outlined here demonstrates how legacy soil datasets can be efficiently leveraged through resampling designs to meet practical constraints while preserving scientific rigour. These findings have broad implications for long-term SOC monitoring, particularly in contexts where site access, budget, and temporal sensitivity pose challenges.
S. Karunaratne, Liz Stower, S. Ugbaje et al.· Soil Research· 1 citation
Soil organic matter (SOM) is a crucial indicator of soil quality and carbon sequestration capacity. In regions with complex topographic and climatic gradients such as Anhui Province, the spatial patterns of SOM and its environmental controls remain insufficiently characterized. This study collected 151 soil samples in 2010 across Anhui Province and derived environmental covariates using GIS and RS technologies. Random Forest, Gradient Boosting, Extra Trees, and LightGBM models were constructed and compared. The Extra Trees model achieved the best performance, with a test-set R² of 0.79, RMSE of 4.55 g/kg, and MAE of 3.66 g/kg. In four-block spatial cross-validation, the R² decreased to 0.60, indicating moderate but limited spatial transferability. The predicted SOM pattern was higher in the southern mountainous region and lower in the northern plains. Nitrogen had the largest model contribution (56.8%). Mean annual precipitation was positively associated with SOM below 1400 mm, and higher predicted SOM occurred at soil pH values of 5.5-7.0. Topographic variables showed terrain-dependent associations, with positive effects in mountainous areas and weaker or negative effects in plains. The calibrated model was applied to environmental covariates representing 2010 and 2024 to produce a 2010 spatial prediction and an exploratory 2024 projection. The projection indicated a small increase in mean SOM from 22.34 to 23.90 g/kg. Because independent SOM observations were unavailable for 2024, the projection should be treated as a covariate-driven scenario rather than a validated estimate. Integrating SHAP with spatial analysis helped identify the modeled environmental controls on SOM variability and may inform soil management in topographically heterogeneous areas.
