Extended Abstract



Introduction



Soil, particularly its organic carbon (SOC), plays a fundamental role in the global carbon cycle and the stability of terrestrial ecosystems. SOC, as the largest terrestrial carbon pool, contributes significantly to improving soil structure, fertility, microbial activity, and water holding capacity, and is considered one of the key indicators of soil sustainability. Accurate assessment of the spatial distribution of SOC is essential for soil resource management, climate change mitigation, and the development of sustainable agriculture. Studies have shown that soil properties, vegetation cover, topographic factors, and geomorphological processes are among the most important controls of SOC spatial variability, influencing carbon accumulation and distribution through the creation of diverse microenvironments.

Recent advances in remote sensing, digital elevation models, and machine learning algorithms have enabled more accurate and higher resolution modelling and spatial estimation of SOC distribution. The integration of these datasets with data mining and geostatistical approaches has significantly enhanced the ability to identify non linear relationships between environmental variables and soil carbon. In semi arid and mountainous regions, including Fereydan, understanding these spatial patterns is crucial for sustainable soil management, as traditional sampling methods are limited by high cost and restricted spatial coverage and therefore cannot provide a complete representation of spatial variability. Accordingly, the aim of this study is to compare the performance of three machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), and K Nearest Neighbors (KNN)—in estimating and spatially mapping SOC and SIC in the Fereydan region.



Methodology





In the spring of 2023, a total of 100 topsoil samples were collected from a depth of 0–20 cm using a stratified random sampling approach. Each sample was composed of four subsamples taken from a 1×1 m area. Soil organic carbon (SOC) was measured using the Walkley–Black method, while soil inorganic carbon (SIC) was calculated by converting calcium carbonate equivalent (CCE) using the relationship SIC (%) = CCE (%) × 0.12.

The research procedure consisted of four main steps: sampling, preparation of environmental variables, selection of influential predictors, and modelling using Support Vector Machine (SVM), Random Forest (RF), and K Nearest Neighbors (KNN) algorithms. Remote sensing data from Sentinel 2 imagery and a 12.5 m resolution digital elevation model were used, and topographic processing was conducted in SAGA GIS. To reduce multicollinearity, a variance inflation factor (VIF < 5) was applied, resulting in the selection of 16 topographic variables and 11 remote sensing indices as final predictors.

Model development was performed using 10 fold cross validation, and model performance was evaluated using R², RMSE, and MAE to ensure a fair comparison among the models. These algorithms were then applied for estimating and spatially mapping the distribution of SOC and SIC in the Fereydan region.



Results and Discussion





The results indicated that for SOC spatial estimation, the K Nearest Neighbors (KNN) model achieved the best performance with R² = 0.91, RMSE = 0.12, and MAE = 0.11. The most influential variables included vegetation cover, valley depth, relative position, and flow convergence index.

For SIC spatial estimation, the Random Forest (RF) model showed superior performance with R² = 0.86, RMSE = 0.98, and MAE = 0.86. The key contributing variables were valley depth, stream network base level, Sentinel 2 band 8, and slope aspect.

SOC was found to be more strongly influenced by vegetation cover and micro topographic factors, whereas SIC was mainly controlled by geological and hydrological conditions. This highlights the different controlling processes governing the two forms of soil carbon.

In terms of spatial distribution, the KNN model showed that SOC was highest in alluvial plains and agricultural lands and lowest in steep and degraded areas, reflecting the effects of erosion, land use, and vegetation cover. In contrast, SIC exhibited its highest values in calcareous formations and high elevation areas, indicating the dominant role of parent material and geochemical processes in arid and semi arid environments.

Overall, the results demonstrated that machine learning models are highly effective in spatial modelling and estimation of soil carbon; however, their performance depends on the type of target variable and underlying environmental conditions.

environmental controls.





Conclusion





Overall, machine learning algorithms demonstrated strong capability in spatial modelling and spatial estimation of SOC and SIC in the Fereydan watershed. The K Nearest Neighbors (KNN) model achieved the best performance for SOC spatial estimation with R² = 0.91, while the Random Forest (RF) model was the most suitable for SIC spatial estimation with R² = 0.86. These results indicate that model selection is dependent on environmental characteristics and the type of soil carbon.

SOC was found to be more strongly influenced by vegetation cover (EVI) and micro topographic factors such as valley depth, flow convergence index, and slope position, whereas SIC was mainly controlled by geomorphological and geological factors including valley depth, drainage network, Sentinel 2 band 8, and slope aspect.

The spatial distribution of SOC and SIC was non uniform. SOC showed its highest values in alluvial plains and agricultural lands and the lowest values in steep terrains, while SIC was more concentrated in calcareous units and high elevation areas. This difference reflects the distinct biophysical and geological controls governing these two forms of soil carbon.

Finally, the results demonstrated that machine learning models are effective tools for spatial assessment and monitoring of soil carbon; however, their accuracy is influenced by limitations such as small sample size and constraints in environmental data availability. The use of more detailed datasets and hybrid modelling approaches is recommended for future studies