Response Surface Sampling of Remotely Sensed Imagery for Precision Agriculture

Abstract

Estimating biophysical characteristics of land surfaces using imagery or spatially contiguous datasets derived from proximal sensors requires developing statistically robust predictive models between the spatial data and ground features. These spatial datasets can contain thousands to millions of points or pixels, so determining the number and location of ground samples to calibrate models is critical. Zoning and co-kriging approaches have been used to identify sampling locations, but they suffer from either ambiguous location identification or the need for 60 or more ground points to produce robust models. In this study, the software ‘EC_e Sampling, Assessment, and Prediction’ (ESAP) was used to select sampling locations from high-resolution (4 m) imagery of a barley crop to develop a predictive regression model for biomass. A normalised difference vegetation index was derived from imagery of an 80 ha field near Rupanyup, Victoria, Australia in 2006. The ESAP software was originally developed to calibrate apparent soil electrical conductivity data to model soil salinity based on response surface theory, but any geo-located spatial dataset can be input. Results showed that the 12 ESAP-selected points produced a statistically significant regression model unbiased by spatial autocorrelation and was better able to predict biomass than models derived from 12 random points and a pooled model derived from 24 sample locations. The correlations between the selected pixels in the imagery and biomass (r ² = 0.38) were not as high as from ground spectral data (r ² = 0.79) collected coincident with biomass sampling, possibly due to misalignment between the imagery and ground data. Digital output from this approach could be used to map soil or plant properties, schedule variable rate applications of chemicals, or be used as inputs to soil and crop simulation models for more accurate site-specific modelling.