Can environmental variables, high sampling density and machine learning deliver detailed maps of soil organic carbon and carbon stock in tropical regions?

Abstract

Studies on the spatial variability of soil organic carbon (SOC) and carbon stocks (CS) at detailed scales are scarce and are mostly conducted in temperate regions, requiring efforts to map SOC and CS with finer resolutions in tropical regions. This work aimed to (1) characterize the spatial distribution of SOC and CS under different land uses; (2) predict SOC and CS using environmental variables and proximal and remote sensing, testing 22 combinations of covariate sets; and (3) map SOC and CS at detailed scale (1 sample per 3.5 ha) and 12.5 m of resolution. 236 soil samples were collected: 180 in a regular grid (200 x 200 m) in two depths (0–20 and 80–100 cm) for model calibration and 56 for validation. Samples underwent SOC and texture analyses, and were scanned by proximal sensors (portable X-ray fluorescence – pXRF, and magnetic susceptibility – MS). Terrain attributes (TA) were calculated from a digital elevation model (12.5 m) and the normalized difference vegetation index (NDVI) and insolation indices were obtained from Sentinel-2 and Landsat 8 imagery. The Random Forest (RF) algorithm was used for modeling. The highest SOC contents were found in an Organossolo (Histosol) under pasture followed by soils under native forest and pasture. High variation of SOC and CS was found even within the same land use, soil class and parent material. The best prediction achieved an R² = 0.47 (SOC at 0–20 cm). Subsurface (80–100 cm) predictions were less accurate. Soil texture, remotely sensed data (NDVI, TA and radiation) and parent material were the most important predictors. High sample density along with environmental variables provided satisfactory accuracy (R² = 0.47). However, since high SOC and CS spatial variation was found, future works are encouraged to obtain more accurate SOC and CS maps at high resolution, e.g., adopting other variables to enhance the results.