CGRG Bibliography of Canadian Geomorphology
Search Results
Author : Coulibaly, P.; and Dibike, Y.
Date : 2004.
Title : Downscaling of Global Climate Model output with dynamic artificial neural networks.
Publication : Eos Transactions. Joint Assembly of the CGU, AGU, SEG and EEGS, Montreal, Canada, May 17-21, 2004.
Issue : 85(17):
Page(s) : H53A-04.
Abstract
The issues of downscaling the outputs of global circulation model (GCM) to a scale appropriate to hydrological impact studies are investigated using functionally different methods. Three types of dynamic artificial neural networks (DANN) with different inherent representations of temporal information are investigated. Time lagged feedforward neural network (TLFN), and two types of globally recurrent neural networks (Elman and Jordan networks) are proposed for downscaling daily precipitation and temperature series for the Serpent watershed in northern Quebec (Canada). The performance of the optimal DANN model is compared to benchmarks from a statistical downscaling model and a stochastic weather generator. Overall, the downscaling results for the current period (1961-2000) suggest that the TLFN is the most efficient of the DANN models tested for downscaling both daily precipitation as well as daily temperature series. The Elman and Jordan networks performed poorly on precipitation downscaling as compared to the TDNN. Furthermore, the different model test results indicate that the optimal DANN model significantly outperforms the statistical and stochastic models for the downscaling of precipitation whatever the season. However, for minimum and maximum temperature, the TLFN and the statistical model are almost equivalent because the inherent physical process is likely less nonlinear. While changes in precipitation between the current and the future scenarios produced by the TLFN are smaller than those produced by the statistical model (except for the winter), they remain significantly larger than those suggested by the stochastic model. Thus suggesting that the TLFN can be a good trade-off alternative to the other models. Changes in streamflows between current and future periods (2020s, 2050s, and 2080s) are also compared and discussed with regard to the downscaling methods.
Bibliography of Canadian Geomorphology