Developing methods to improve sampling efficiency for automated soil mapping

Abstract

The goal of this project was to develop methods to sample spatial variables, such as soil or crop properties, which are efficient and cost-effective despite the fact that we start with little or no information about the spatial variability of the variable. Commonly when we begin a survey we do not know in advance how intensively to sample, and so we run the risk of completing then survey then finding that we have substantially over-sampled and so wasted effort, or that we are not able to produce a reasonable map from the our data because they are too sparse.

We developed an approach to optimization of sampling for a single-phase geostatistical survey, showing how the combined uncertainty of our model of spatial variation (variogram), and predictions under spatial variation could be quantified and minimized, assuming some prior distribution of variogram parameters. We showed, using simulation, how this scheme successfully models prediction error variance, and how under different underlying kinds of spatial variation the resulting sample schemes achieve the dual goal of allowing adequate estimates of the variogram and disposing sample points from which to map the variable.

We then developed a fully adaptive sample scheme. Under this the sampling is divided into phases. In the initial phases (reconnaissance) our uncertainty about the required final sample intensity is reduced by sampling designed to yield maximum information, and at the end of each phase we are presented with a representation of this uncertainty which allows us to decide whether further reconnaissance sampling is justified or whether we should proceed to a final sampling phase before mapping. This final phase may be designed so as to allow for the existing observations, and so to save on total sampling effort. We demonstrated the value of this method by simulation.

We built a field system to implement these algorithms for mapping the water content of the soil with a sensor. This uses GPS to record locations, and to guide the user to selected sample sites. We applied the system in a survey of a field, and showed how it identified the relatively sparse sample effort needed to map the variable to a specified precision.

In conclusion, we have shown how substantial efficiencies in sampling are possible using appropriate algorithms. Some of the methods we have used, notably the spatial simulated annealing algorithm to select sample points, could be applied to simpler sampling problems that farmers and agronomists face.