Predicting spatial variability of selected soil properties using digital soil mapping in a rainfed vineyard of central Chile

Abstract Soil physical properties influence vineyard behavior, therefore the knowledge of their spatial variability is essential for making vineyard management decisions. This study aimed to model and map selected soil properties by means of knowledge-based digital soil mapping approach. We used a Random Forest (RF) algorithm to link environmental covariates derived from a LiDAR flight and satellite spectral information, describing soil forming factors and ten selected soil properties (particle size distribution, bulk density, dispersion ratio, Ksat, field capacity, permanent wilting point, fast drainage pores and slow drainage pores) at three depth intervals, namely 0–20, 20–40, and 40–60 cm at a systematic grid (60 × 60 m2). The descriptive statistics showed low to very high variability within the field. RF model of particle size distribution, and bulk density performed well, although the models could not reliably predict saturated hydraulic conductivity. There was a better prediction performance (based on 34% model validation) in the upper depth intervals than the lower depth intervals (e.g., R2 of 0.66; nRMSE of 27.5% for clay content at 0–20 cm and R2 of 0.51; nRMSE of 16% at 40–60 cm). There was a better prediction performance in the lower depth intervals than the upper depth intervals (e.g., R2 of 0.49; nRMSE of 23% for dispersion ratio at 0–20 cm and R2 of 0.81; nRMSE of 30% at 40–60 cm). RF model overestimated areas with low values and underestimated areas with high values. Further analysis suggested that Topographic position Index, Topographic Wetness Index, aspect, slope length factor, modified catchment area, catchment slope, and longitudinal curvature were the dominant environmental covariates influencing prediction of soil properties.