Investigation of Sample Numbers Needed to Map Spatial Changes in Soil Moisture Using Random Forests and Z-Score Calibration for Precision Irrigation of Turfgrass

In the USA, agriculture is the largest consumer of freshwater resources, and precision irrigation (PI) can conserve water significantly while maintaining crop yield. Current approaches to soil volumetric water content (VWC) mapping for PI rely on installing a costly soil moisture sensor within each of 4–5 management zones per field. Although this strategy provides temporally dense data, it is spatially sparse. Alternatively, spatially dense remotely sensed data require calibration with in situ soil moisture measurements, which are expensive and labor intensive to obtain. Previous research indicates that soil VWC zones must be regularly reassessed, a process that is impractical without low-cost soil VWC sensors. In anticipation of deploying dense networks of inexpensive soil moisture sensors for PI in large turfgrass fields, we investigate the mapping errors and optimal sampling density required for accurate soil VWC mapping using random forests (RFs) and z-score calibration in two turfgrass sports fields in Utah. Dense sampling of soil VWC was undertaken at 101 and 103 points in each field with a theta probe. These data were systematically sub-sampled to quantify errors in z-score soil moisture maps generated with varying sample sizes. A jack-knife procedure was employed to determine the optimum number of sensors required to produce accurate RF-based soil moisture maps. The RF approach also allows identification of the most influential covariates for soil VWC prediction. For RFs, 21–79 samples were needed to characterize changing spatial patterns in fields with mean absolute errors (MAEs) of 1.39–9.71%, but for most dates only 25–40 samples were needed. The z-score calibration produced MAEs of 1.38–10.44% with as few as 10–15 samples, but the spatial patterns remain static and only the magnitude of values changes. Therefore, using RFs with 40–60 sensors was recommended to allow for accurate mapping despite dropped signals and broken sensors.