STATISTICAL AND NEURAL METHODS FOR SITESPECIFIC YIELD PREDICTION

Abstract

Understanding the relationships between yield and soil properties and topographic characteristics is of criticalimportance in precision agriculture. A necessary first step is to identify techniques to reliably quantify the relationshipsbetween soil and topographic characteristics and crop yield. Stepwise multiple linear regression (SMLR), projection pursuitregression (PPR), and several types of supervised feedforward neural networks were investigated in an attempt to identifymethods able to relate soil properties and grain yields on a pointbypoint basis within ten individual siteyears. To avoidoverfitting, evaluations were based on predictive ability using a 5fold crossvalidation technique. The neural techniquesconsistently outperformed both SMLR and PPR and provided minimal prediction errors in every siteyear. However, insiteyears with relatively fewer observations and in siteyears where a single, overriding factor was not apparent, theimprovements achieved by neural networks over both SMLR and PPR were small. A second phase of the experiment involvedestimation of crop yield across multiple siteyears by including climatological data. The ten siteyears of data were appendedwith climatological variables, and prediction errors were computed. The results showed that significant overfitting hadoccurred and indicated that a much larger number of climatologically unique siteyears would be required in this type ofanalysis.