Comparaison de techniques statistiques utilisées dans l'élaboration de modéles prévisionnels phénoclimatiques

Abstract This study was undertaken to compare the use of two statistical approaches, the coefficient of variation and the standard error of prediction, for the determination of a threshold temperature of development and a starting date of summation in order to obtain the least variable degree-day models. Five phenophases observed on lilac and honeysuckle cultivars were analysed at five locations in the southern part of Quebec; seven years of data were available at each site. Three starting dates were used: January 1, March 1 and April 1. The standard error of prediction is easier to interpret than the coefficient of variation and is especially useful for comparing the forecasting efficiency of models. However, for the identification of parameters essential for the determination of the thermal constant, both approaches were equivalent. January 1 and March 1 results yielded the same results. However, January 1 seems more appropriate for predicting phenophases related to leafing and March 1 to flowering ones. Both dates provided approximately the same threshold temperatures, which were -1°C at the occurrence of the first leaf, O°C at the time when all the buds had leafed, 2°C at the opening of the first bloom, 4°C at the peak of full bloom and 5°C at the end of bloom. For each phenophase April 1 was too late to start the summations. Phenophases related to flowering were better adjusted to linear models than those of the leafing period. Our results illustrate that a degree-day model may be applied over a relatively large area, showing some differences in the thermal regime, without requiring any adjustment to the basic parameters of the constant. We also stress the fact that the efficiency of these models is not necessarily linked to a unique choice of base temperature and starting date. These findings may be of interest to rationalize data collection in agroclimatic forecasting networks.