Using targeted sampling to process multivariate soil sensing data

Most soil properties sensed on-the-go (e.g., electrical conductivity, capacitance, optical reflectance, mechanical resistance and soluble ion activity) are not directly related to the agronomic parameters used to make management decisions. Nonetheless, these sensors provide an opportunity to obtain fine-resolution data about the spatial variability of soil in agricultural fields, rapidly and at a relatively low cost. To process this information, a limited number of targeted samples must be collected and undergo conventional laboratory testing for site-specific calibration of the sensor data. Selecting sampling locations based on multiple sensor data layers is an important process and, in practice, is conducted in a very subjective manner. This paper discusses an analytical methodology to assess the quality of targeted sampling strategies for on-the-go soil sensor data calibration prior to site-specific soil treatments, and demonstrates the potential for the automated selection of sampling sites. The methodology uses an arbitrary objective function that maximizes the spread among sensor output, local homogeneity (spatial uniformity around each location), and physical coverage across an entire field. Soil pH and electrical conductivity maps of a 23-ha agricultural field were used to illustrate the applicability of this method. From those considered, a Latin hypercube sampling (LHS) procedure with homogeneity and field coverage constraints provided the highest probability of maximum objective function outcomes, when individual criteria were normalized by the median of a large number of random prescription sets.

Research Highlights
► Sampling locations to calibrate an on-the-go soil sensor are selected subjectively.
► An objective function was developed to evaluate the quality of such locations.
► Data and space coverage as well as local homogeneity are three criteria combined.
► Different sampling strategies have been used to explore sensitivity.
► Constrained categorical stratification and Latin hypercube sampling were the best.