Regional prediction of soil organic carbon dynamics for intensive farmland in the hot arid climate of India using the machine learning model.

To manage soil and agriculture sustainably, it is crucial to keep updated on changes in the pools of soil organic carbon (SOC). Using tools from digital soil mapping (DSM), this case study simulates the dynamics of total and active organic carbon. The aim of present study was to assess the performance of various machine learning techniques for SOC fractions. The SRTM DEM and its derivatives were used to predict total organic carbon (TOC) and its fractions for a 4000 ha study area in Central State Farm, Sardargragh, Rajasthan. For prediction, Extreme Gradient Boosting (XG Boost), Cubist, and Random Forest (RF) models were used. For modeling purposes, the SOC fractions like very labile C (VLC), labile C (LC), less labile C (LLC) was analyzed. The range of the WBC content was 0.97 to 6.85 g kg^-1. Both the calibration and validation sets of VLC showed outstanding results from the RF model (R²c = 0.948, RMSEc = 0.174 g kg^-1and R²v = 0.204, RMSEv = 0.213 g kg^-1). The XG Boost model, on the other hand, had the next lowest accuracy and explained just around 10.06–36.8% of the variation of C fractions. The cubist model performed the worst in both the calibration and validation sets. Clay is the most significant predictor of the TOC, WBC, LLC, NLC and PC. The predicted SOC varied from 3.06 to 9.41 g kg^-1, 0.34 to 1.42 g kg^-1, 0.73 to 2.45 g kg^-1, 0.28 to 3.30 g kg^-1 and 1.05 to 4.12 g kg^-1 in TOC, VLC, LC, LLC and NLC, respectively. RF proved to be the most accurate and least uncertain model when it came to predicting regional SOC. Based on the findings of our simulation; it appears that SOC may be approximated rather accurately and with ease. In general, stakeholders, decision-makers, and applicants in agricultural management approaches toward precision agriculture may find their high-resolution maps helpful. [ABSTRACT FROM AUTHOR]