A neural network experiment on the site-specific simulation of potato tuber growth in Eastern Canada

The objective of this work was to optimize a neural network (NN) for modelling potato tuber growth and its in-field variations in eastern Canada. In addition to climatic inputs, the cumulative and maximal leaf area index (LAI) were incorporated to account for in-field scale variability. Soil and genetic parameters were assumed to be integrated in LAI as suggested by earlier work. Each input and combination of inputs was evaluated from the changes they induced in MAE (mean absolute error) and RMSE (root mean square error). Results using data from several replicated on-farm experiments between 2005 and 2008 suggest that a NN model using cumulative solar radiation, cumulative rainfall and cumulative LAI can adequately model site-specific tuber growth. The MAE of the retained model was 209 kg DM ha−1, which represents less than 4% of the mean final tuber yield for the 3 years of the study. Non-linear effects of explicative variables on tuber yield were attested by comparing the results of the NN simulations to those of a multiple linear regression (MLR). The failure of MLR to simulate temporal discontinuities in tuber growth supports the use of a non-linear approach such as a NN to model tuber growth.