Spatial statistical analysis of soil properties and crop yields for precision agriculture applications

Abstract

Soil properties, as well as many natural resource properties, vary continuously in space. Geostatistical techniques attermt to quantify and predict the variation of these spatial properties. Geostatistics and the theory of "regionalized variables" utilizes the fact that properties in space tend to display spatial structure and spatial dependence. Once the underlying spatial dependence is quantified, geostatistical techniques are used to interpolate or estimate a continuous 'surface' of the property of interest. Recent advances in global positioning systems and geographic information systems for agriculture have resulted in the rapid development of research initiatives in the area of spatial statistics applied to agricultural systems with the ultimate goal of implementing precision farming technologies. The objective of this study was to examine several spatial sampling and spatial interpolation techniques related to precision farming technologies. This dissertation is composed of five chapters. The common theme throughout this study is the necessity of capturing the spatial dependence of the soil parameter of interest. Capturing spatial dependence is the key to properly interpolating soil properties and ultimately to improving precision farming technologies. Chapter 1 is a general introduction to spatial interpolation techniques for natural resources data. The objective of chapter 2 was to evaluate an alternative procedure for principal component kriging designed to benefit from the utilization of a spatial correlation matrix. Principle component kriging has been purported to take advantage of the orthogonality of principal components. The assumed benefit of principal component kriging is that computational effort is reduced by kriging n principal components as opposed to cokriging n variables. My hypothesis was that principal components constructed with a spatial correlation matrix are more effective than 'linear' principal components for kriging purposes because kriging is an inherently spatial interpolation technique. The spatial correlation matrix was constructed with a modified Moran's I statistic. The goodness-of-fit statistics ranged from 0.771 to 0.965 and correlations from back-transformation ranged from 0.790 to 0.986. Based on the goodness-of-fit statistic and back-transformation results, both techniques were equally effective for principal component kriging. The objective of chapter 3 was to examine an alternative spatial analysis tool called the cumulative correlogram that is useful for understanding the underlying spatial patterns of soil properties. My hypothesis was that cumulative correlograms would be more useful than standard correlograms or variograms because cumulative correlograms can elucidate the maximum scale-of-pattern of a soil parameter. In most soil sampling situations, the scales-of-pattern associated with soil parameters occur at separation distances much smaller than the separation distances observed with a typical grid sampling approach. Kriging models for grid-sampled soil parameters performed poorly based on goodness-of-fit statistics. Evaluation of point autocorrelation coefficients and cumulative correlograms indicated that two factors contribute to the poor performance of the kriging models: anomalous characteristics present in the grid-sampled data sets, and, large separation distances among grid samples. The objective of chapter 4 was to develop an alternative sampling approach designed to capture small scale variability of soil parameters. My hypothesis was that utilizing a cluster sampling approach as well as auxiliary data layers would be a more effective sampling approach because small scale variation would be present data sets. The design utilizes auxiliary data layers and an alternative cluster-sampling approach to data collection. The new, alternative sampling design is compared to a traditional grid-sampling design. The cluster-sampling design improved all parameter estimates; the grid-sampling design improved one parameter. Average improvement for the cluster-sampling was 34% (minimum = 1%, and maximum =72%). Bias estimates from the cluster-sampling design were small and estimates of root mean squared error suggest the cluster-sampling design is a better predictor of small-scale variation when compared to the grid-sampling design. Recent research has focused on delineating field-specific soil-productivity management zones to be used as prescription, fertilizer application maps for variable rate application technologies. The objective of chapter 5 was to analyze four techniques for delineating soil-productivity management zones. Each of the methods uses a unique set of soils, yield, and or remotely sensed data. My hypothesis was that management zone techniques that utilize soils data at the outset of the delineation procedure would be more effective than techniques that fail to utilize soils data. Analysis of variance for the majority of yields among management zones indicated yields among management zones were different. A non-parametric analysis of crop yields also provided evidence to conclude that management zone delineation techniques resulted in yield patterns that were different from random yield patterns. Overall, delineation techniques that combined secondary soils information and soil-sample analysis results were the most effective techniques.