Artificial neural network configurations for predicting corn yield as a function of water regime

Several feedforward artificial neural networks with error back-propagation were trained to predict corn yield as a function of the timing and volume of water applications. One set of artificial neural networks (ANNs) was developed using simulated water delivery and yield data. The growing season was divided into time intervals of equal length, and each ANN input node represented the total amount of water applied within each time interval. When an optimal time interval was found and sufficient training data were provided, independent tests of the trained ANNs gave r{dollar}sp2{dollar} correlations greater than 0.90 between the predicted and actual yields. These correlations tended to approach a maximum value asymptotically as the size of the training data set increased.; In a similar fashion, another set of ANNs was developed using six years of field data from North Platte, Nebraska. An additional input node incorporated a measurement of soil moisture taken early in the growing season. Depending on which data were selected for training and testing, r{dollar}sp2{dollar} correlations between the predicted and actual relative yields were as high as 0.89. An ANN analysis of crop sensitivity to drought as a function of growth stage produced values consistent with data in the literature.; An ANN crop production model can be used with three different iterative algorithms to allocate water among growth stages in a manner which is optimal or nearly optimal. The first of these algorithms begins with a superfluous amount of water allocated across the growing season and, with each iteration, decreases the water allocation to that stage in which the decrement will result in the least reduction in yield. The second of these algorithms begins with no water allocated across the growing season and, with each iteration, increases the water allocation to that stage in which the increment will result in the greatest return to yield. The third of these algorithms begins with an assumed seasonal supply of water applied at a constant rate across the growing season and, with each iteration, reallocates a marginal amount of water from a less productive stage to a more productive stage.