A Comparison of Four Spatial Regression Models for Yield Monitor Data: A Case Study from Argentina

Abstract

The gap between data analysis and site-specific recommendations has been identified as one of the key constraints on widespread adoption of precision agriculture technology. This disparity is in part due to the fact that analytical techniques available to understand crop GIS layers have lagged behind development of data gathering and storage technologies. Yield monitor, sensor and other spatially dense agronomic data is often autocorrelated, and this dependence among neighboring observations violates the assumptions of classical statistical analysis. Consequently, reliability of estimates may be compromised. Spatial regression analysis is one way to more fully exploit the information contained in spatially dense data. Spatial regression techniques can also adjust for bias and inefficiency caused by spatial autocorrelation. The objective of this paper is to compare four spatial regression methods that explicitly incorporate spatial correlation in the economic analysis of variable rate technology: (1) a regression approach adopted from the spatial econometric literature; (2) a polynomial trend regression approach; (3) a classical nearest neighbor analysis; and (4) a geostatistical approach. The data used in the analysis is from a variable rate nitrogen trial in the Cordoba Province, Argentina, 1999. The spatial regression approaches offered stronger statistical evidence of spatial heterogeneity of corn yield response to nitrogen than ordinary least squares. The spatial econometric analysis can be implemented on relatively small data sets that do not have enough observations for estimation of the semivariogram required by geostatistics. The nearest neighbor and polynomial trend analyses can be implemented with ordinary least squares routines that are available in GIS software. The main result of this study is that conclusions drawn from marginal analyses of this variable rate nitrogen trial were similar for each of the spatial regression models, although the assumptions about spatial process in each model are quite different.