Extension of crop model outputs over the land surface by the application of statistical and neural network techniques to topographical and satellite data

The use of crop simulation models to evaluate cultivars and cropping practices has developed greatly in the last few years. These tools can provide unique advantages in several situations, for example, allowing a quick response when new needs arise or to extrapolate results of field experiments in different environmental (climate, soil) and agronomic (cultivars, cropping systems) situations. The operational utilisation of the results of models is however bounded by the problem of extrapolating then to all points on the land surface, which is not always a trivial task in topographically complex regions. The present work investigates the use of different methodologies for the extension of the outputs of a grapevine model in a rugged region of central Italy, Tuscany. In particular, 2 approaches were considered, the first based on statistical assumptions and the second on neural network reasoning. These techniques were applied using, as input parameters, topographical information layers and low-resolution satellite data related to vegetation development. The results obtained show that, in general, the neural network approach produced higher accuracy levels than the statistical approach, but the latter was more capable of merging information coming from different sources. Moreover, the estimates derived from the 2 methods show different spatial patterns and ranges, which must be taken into account when considering these approaches for possible operational uses.