STOCHASTIC IMAGING OF SOIL PARAMETERS TO ASSESS VARIABILITY AND UNCERTAINTY OF CROP YIELD ESTIMATES

Modern site-specific agriculture uses yield estimates based on estimat es of soil properties in locations other than the sampling points. Tec hniques are needed to assess the uncertainty of these soil property es timates. Such uncertainty assessments can be based on stochastic imagi ng of soil parameters: a technique that consists of generating many eq uiprobable maps oi the parameters for the same site. The objective of this study was to use stochastic imaging of the available soil water c apacity (AWC) and a soybean crop model GLYCIM to simulate variability and uncertainty in crop yield estimates as related to soil sampling de nsity and weather patterns. First, we generated an exhaustive AWC data set on a fine grid, simulated yields at the fine grid nodes, and cons idered the results as the 'true' yield values. Then we sampled the exh austive data set using sparse grids, and carried out stochastic imagin g of AWC using genetic algorithms. We simulated yields for each image, and calculated the errors in yield estimates as the differences betwe en the 'true' yields and yields from the images. The probability distr ibutions of the errors were used to quantify the uncertainty. The fine grid for the exhaustive dataset was 25 X 25 m. The sparse grids at 50 x 50 m and 100 x 100 m corresponded to typical research and commercia l soil sampling densities, respectively. The simulations were repeated for three different weather patterns. Results showed that the distrib utions of errors in yield estimates were affected by weather pattern, and the temporal variability in yield error estimates could not be ove rridden by improvements in spatial variability estimates at the sparse sampling densities that we considered. A nonlinearity in the yield re sponse to AWC caused increases in the probabilities of obtaining very small and very large errors in yield estimates. Stochastic imaging of soil properties is a desirable step to assess the efficiency of a part icular sampling density to be used repeatedly over several years. (C) 1998 Elsevier Science B.V. All rights reserved.