Soil carbon prediction under different land uses by integration of remote sensing and machine learning algorithms in a semi-arid watershed, Iran

Soil organic carbon content (SOCC) plays a vital role in restoring soil health and the sequestration of atmospheric carbon which is responsible for warming the atmosphere. Hence, there is a need to model and predict SOCC across different land-use types in semi-arid conditions. In this work, we model and predict SOCC for land-use types such as farmlands, rangelands, and bare lands in a semi-arid watershed in northwest Iran. SOCC was measured at 106 locations spanning different land-use types. MLR (multiple linear regression), ANN (artificial neural network), and SVM (support vector machine) models were used to estimate SOCC, using input variables including topographic factors and remote sensing indices. Predictive models were generated for each individual land-use type as well as all land-use types combined (full dataset). The results showed that the ANN model was best for predicting SOCC in the combined dataset (ME = 0.045%, RMSE = 0.269%, and R2 = 0.720). For the individual models for each land-use type, the ANN model was also shown to perform best for predicting SOCC, with results for farmlands (ME =  − 0.035%, RMSE = 0.287%, and R2 = 0.735), for rangelands (ME = 0.019%, RMSE = 0.236%, and R2 = 0.634) and for bare lands (ME =  − 0.064%, RMSE = 0.174%, and R2 = 0.443). It was concluded that by using an ANN and integrating topographic factors and remote sensing indices, it was possible to model and predict SOCC accurately in semi-arid areas of northwest Iran.