Predicting the spatial distribution of organic carbon in soil by combining machine learning algorithms and spline depth function in a part of Golestan Province, Iran

Abstract

As a critical element of soil quality, soil organic carbon (SOC) constitutes nearly 75 % of the active carbon content in terrestrial ecosystems and is essential for agricultural productivity. Estimation of soil organic carbon is one of the requirements of effective soil management planning. The present paper aimed to test the performance of Cubist (Cu), Quantile Regression Forest (QRF), and Random Forest (RF) methods in predicting the distribution of SOC content at four soil depths of 0–15, 15–30, 30–60 and 60–100 cm in Golestan Province of Iran. To achieve this goal, the mentioned models were trained with 105 soil profiles across the study area using environmental covariates which had been obtained from DEM, rain and piezometric maps, and remote sensing data extracted from Landsat 7 ETM⁺. Results revealed that the mean SOC values varied between 0.39 % and 1.24 %. All three predictive models had the highest performance in predicting SOC at 15–30 cm depth (R² = 0.45, RMSE = 0.34, and MAE = 0.25). Although the predictive models were similar in terms of validation metrics criteria (R², RMSE, and MAE), the QRF predictions based on the Taylor diagram had a higher agreement with the measured SOC distribution, resulting in the identification of QRF as the leading model in performance. Results of the present paper indicate the high potential of rainfall, piezometric, MRVBF, MRRTF, and valley depth to predict SOC concentration distributions. These findings will contribute to further research on SOC prediction models.