PREDICTING CORN GRAIN YIELD SPATIAL PATTERN: COMPARISON OF TECHNIQUES

摘要

A component that has been missing in most site-specific management efforts is a strong link between the diverse layers of data describing existing variability and the consequent spatially variable management recommendation. An accurate method that relates resource variability and final grain yield is, therefore needed. In this paper we compare three procedures: regression methods, neural networks and crop simulation modeling. The methods have different degrees of empiricism as well as advantages anddisadvantages. The properties of each method are discussed in the paper. We use data from a cornfield in Michigan collected in two successive years. The major variables influencing yield are plant available soil water and plant population. Since we havetwo years of data we are able to calculate the root mean square error of description as well as the root mean square error of prediction. The regression methods were totally inapplicable in the case study. The main reason was the dynamic aspect of soil water stress. Due to the complex interactions and dynamic component of this water stress, crop simulation model out-performed all neural networks tested.