Advancement of new technologies including high precision GPS, remote sensing and GIS has greatly improved our ability of the spatial prediction of soil properties using digital soil mapping (DSM). DSM is considered highly valuable for planning and implementing agro-ecological projects, or evaluating the availability of key ecosystem services on farmland. Soil mappers are constantly improving and introducing new DSM techniques for enhancing the predictive capability of statistical and geostatistical models; however, field sampling design, which is the primary input to the models, has only recently begun to receive adequate attention. The accuracy of DSM model output is highly reliant on the number of field samples and design of the sample collection. There is a clear trade-off between higher accuracy achieved by more field samples, and costs for labour, time and analysis. Developing a methodology to identify the sampling size that will optimize for both accuracy and cost is a critical component for maximizing DSM capabilities. In this research, a field-scale study was conducted in a crop field in Delta, BC, Canada to evaluate various sampling sizes and designs for DSM of soil properties.
A 54-hectare field was sampled using a 40 x 40 m grid sample locations were recorded with a DGPS system providing ~10 cm of accuracy. The grid sampling design was then revisited with Conditional Latin Hypercube Sampling (CLHS) techniques which selected 20%, 40%, 60% and 80% of the total samples based on the environmental covariates which include drone image derived RGB vegetation index, the field management layout, and a suite of topographic indices derived from DEM. Each of the sample designs was utilized for modeling soil properties (e.g. texture, electrical conductivity, organic matter and total nitrogen) using ordinary kriging, universal kriging and hybrid regression kriging. For each sampling design, 75% of sample points were used for prediction and 25% for accuracy testing. The analysis showed that model performance, as measured by R2 and root mean square error (RMSE), improved with between 20% to 60% of samples but increasing the sample number past 60% had no impact on the prediction capacity of the models. In addition, it was observed that R2 and RMSE of the universal kriging and hybrid regression kriging is 10-25% better than that of ordinary kriging for selected soil properties. These results suggest that optimizing sample design is highly valuable for DSM and may save resources invested in field sampling for agro-ecological projects.