Soil Organic Carbon Stock Estimation Using Legacy Data: A Case Study of North Fluminense Region—BR

This work aimed to model and map the organic carbon stock (SOCS) distribution of soils up to 1 meter (m) depth using legacy data in the northern region of Rio de Janeiro (Brazil). The data belongs to the PROJIR dataset, which was generated in 1983 and covered an area of approximately 30% of the northern region of the State of Rio de Janeiro. The study rescued 161 profiles (121 training and 40 external validation) of soil containing organic carbon and soil density, which allowed the calculation of SOCS up to 1 m of soil depth. Furthermore, the study aims to compare the performance of geostatistical algorithms (Ordinary Kriging—OK) and Random Forest (RF) for SOCS modeling and mapping. The RF algorithms were arranged into three groups of models: (1) All covariates and covariates selected by (2) Recursive Feature Elimination, and (3) Expert Knowledge selection (EK). In all, 21 covariates were evaluated as predictors: 12 derived from the DEM; geomorphological and geological forms; annual mean precipitation and temperature; maps of sand, clay, and silt obtained via OK; and latitude and longitude coordinates. The SOCS ranged from a minimum of 2.18 to a maximum of 104.47 kg C.m−2, with most soils having SOCS of up to 30 kg C.m−2. Among the evaluated approaches to map SOCS, considering only the mean square error (MSE), the root mean squared error (RMSE), and Bias metrics, the RF model 2 stood out, followed by RF model 3 and RF model 1. The OK was the best concerning R2 and mean absolute (MAE). The model 3 (EK) was considered the most consistent once it balanced a good relative improvement in the metrics about the OK map, and best met the plausibility, interpretability, and explainability criteria.