Evaluation of STARDEX indices calculated from linear and nonlinear statistical downscaled daily temperatures

When determining the hydrological impacts of climate change, climate information at the watershed level is required. Global Climate Models data must therefore be downscaled to local meteorological variables. In this study, the Canadian Global Circulation Model version 3.1, and the A1B and A2 simulations from the Special Report on Emissions Scenarios (SRES) were used in conjunction with nonlinear Artificial Neural Networks (ANN) and Multiple Linear Regression to statistically downscale values of maximum and minimum temperature to five weather stations on Vancouver Island, Canada. Linearly and nonlinearly downscaled values from the control period (1960-2000) and the 21st century simulations were evaluated using six indices developed by the European Statistical and Regional Downscaling of Extremes project (STARDEX). The indices evaluated were: 90th percentile of maximum temperature, 10th percentile of minimum temperature, growing season length, number of frost days, heat wave duration, and intra-annual extreme temperature range. The results show that nonlinearity in cross-validated ANN models brings a marginal improvement in the statistical downscaling of daily temperature for the control period. Both downscaling techniques predict an increase of maximum and minimum temperatures, intra-annual extreme temperature range, heat wave duration, and growing season length for both scenarios; as well as a decrease in the number of frost days.