Local calibration of bulk density models for agricultural soils in an inter-Andean valley of the Peruvian Central Highlands

Soil bulk density (BD) is an essential indicator of soil quality, reflecting compaction, porosity, and water and nutrient retention capacity. However, its direct determination is costly and labor-intensive. This study developed and calibrated predictive models for BD in agricultural soils of an inter-Andean valley of the Peruvian Central Highlands using physicochemical properties (pH, electrical conductivity, organic matter, P, K, sand, silt and clay). Sixty-seven soil samples were analyzed using five modeling approaches: multiple linear regression (MLR), partial least squares regression (PLSR), artificial neural networks (ANN), classification and regression trees (CART), and random forest (RF). model performance was evaluated through fivefold cross-validation and goodness-of-fit metrics (coefficient of determination, R2; root mean square error, RMSE; mean absolute error, MAE; and performance to deviation, RPD). The RF model gave the best overall fit among the evaluated models (R2 = 0.621, RMSE = 0.0862 g·cm−3, MAE = 0.0693 g·cm−3, RPD = 1.64), outperforming MLR (R2 = 0.614), PLSR (R2 = 0.608), ANN (R2 = 0.584) and CART (R2 = 0.514). Local calibrations significantly reduced errors compared to standard pedotransfer functions, which typically have standard errors of 0.14–0.20 g·cm–3 and R2 < 0.60. Results demonstrate the potential of local models, particularly machine learning approaches, to generate rapid, low-uncertainty BD estimates supporting soil fertility management and sustainable agricultural planning.