NEURAL NETWORK TO SIMULATE POTATO TUBER YIELD IN EASTERN CANADA

Classical mechanistic crop simulation models are useful tools in agricultural decision-making. Potato growth model such as SUBSTOR, for example, simulate plant development and agronomic yields as a function of weather, soil conditions and crop management. However, in a context of precision farming, the utility of this model might be limited by the difficulty in assessing in-field crop growth variation. Although such models can often be locally calibrated by the use of genetic coefficients, the range of value of these coefficients are often too small to fully cover the variations commonly observed for a specific site. Moreover, they are generally fixed for a given location, making it impossible to adequately assess in-field crop growth variation. In this study, we present a simple neural network for estimating potato growth from LAI measure¬ments and widely available meteorological data. Advantages over the classical mech¬anistic approach include: 1) the capacity to better account for the non-linearity of the phenomenon under study, 2) the fewer field measurements needed, and 3) the spatial dimension given by the use of LAI measurement as principal input.