Random forest-based soil organic carbon prediction using landsat temporal indices and global covariates in a Peruvian coastal valley

This study evaluated whether a global modeled soil product adds predictive value to locally calibrated digital soil mapping of topsoil soil organic carbon (SOC; 0–30 cm) in an irrigated Peruvian coastal valley. Random forest models used 135 composite samples collected in 2024–2025, terrain derivatives, temporally weighted Landsat indices, and OpenLandMap-soildb predictions. Configurations were compared using internal performance assessment with 200 random 80/20 partitions and 50 repetitions of block spatial cross-validation (SpCV); no independent external validation dataset was used. The retained model (elevation, slope, focal valley wetness index, topographic position index, and OpenLandMap-soildb p84) showed weak transferable predictive ability within the valley, with SpCV R² = 0.155 (95% CI 0.077–0.221), compared with 0.189 under random partitioning. Terrain alone achieved SpCV R² = 0.090; adding p84 increased this by 0.065, whereas p84 alone performed poorly (R² = –0.058). OpenLandMap therefore offered limited supplementary information and could not replace local environmental covariates or field observations. Predictions were displayed on a 30 m grid, but p84 retained its effective 120 m support and the composite observations represented management areas up to 3 ha. Quantile regression forest intervals achieved empirical coverage of 0.88 at the nominal 90% level under both internal resampling schemes, with a median mapped width of 1.51% SOC. The area of applicability included 97.5% of mapped pixels but assessed covariate similarity, not prediction reliability. The main contribution is an evaluation protocol showing why single-partition model rankings are unstable; the map supports screening and sampling prioritization, not field-level decisions.