Spatial mapping and predictive modeling of soil organic carbon stocks in Vermont agricultural lands using machine learning and environmental variables

• Developed a novel spatial SOC stock map for Vermont’s agricultural lands. • Applied Random Forest model, outperforming kNN and Cubist for SOC prediction. • Identified climate variables, particularly precipitation, as key drivers of SOC. • Discovered high SOC stocks in northeastern Vermont, influenced by climate and topography. • Results contribute to Vermont’s goals and similar regions worldwide for soil health and climate change mitigation. Understanding and accurately predicting soil organic carbon (SOC) stocks in agricultural lands play a vital role in mitigating climate and sustainable land management. However, existing studies often lack high-resolution SOC stock maps at regional scales, limiting their applicability for site-specific land management. This research provides a novel and comprehensive framework for generating high-resolution (10-m) SOC stock maps of agricultural lands in Vermont, USA—one of the first efforts of its kind in a temperate, data-scarce region—using digital soil mapping (DSM). We compiled 361 topsoil samples (0 to 30 cm depth) and then applied Cubist, kNN (k-Nearest Neighbors), and RF (Random Forest) machine learning (ML) algorithms to predict SOC stocks (t ha −1 ) using environmental variables including climate, terrain, remote sensing, and soil characteristics. Prior to modeling, we used the Boruta algorithm to identify significant variables for SOC stock. Model performance was assessed using cross-validation (70% training and 30% validation). Validation parameters indicated that the RF ML algorithm outperformed others in predictive accuracy. Dynamic variables, including climate and biota, were the most influential variables in defining SOC distribution, while static variables, like terrain attributes and soil properties, were less influential. Spatial prediction maps revealed high SOC stocks in the northeastern part of Vermont, where there is high precipitation and elevation. The study also includes novel spatial uncertainty quantification across different land use types, offering practical insights into prediction confidence and C incentive targeting. Uncertainty in SOC stock prediction accuracy ranged from approximately 3.35 % to 3.63 %, with mean SOC stock values of 94.3 ± 3.42 (t ha −1 ) for crops, 100.0 ± 3.35 (t ha −1 ) for hay, and 96.1 ± 3.30 (t ha −1 ) for pasture. Our findings provide a foundational SOC stock map for Vermont’s agricultural lands, revealing key spatial distribution and drivers. The strong influence of climate variables suggests that adaptation strategies should account for regional climatic conditions, and incentives for soil C sequestration should be location specific. This research enhances the capacity to make soil carbon-informed agricultural management decisions aimed at mitigating the impacts of climate change, while also highlighting the need for further research to overcome the limitations of current SOC mapping approaches in Vermont and similar temperate agricultural regions worldwide.