Downscaling probabilistic seasonal climate forecasts for decision support in agriculture: A comparison of parametric and non-parametric approach

Seasonal climate forecasts (SCF) are produced operationally in tercile-probabilities of the most likely categories, e.g., below-, near- and above-normal rainfall. Inherently, these are difficult to translate into information useful for decision support in agriculture. For example, probabilistic SCF must first be downscaled to daily weather realizations to link with process-based crop models, a tedious process, especially for non-technical users. Here, we present two approaches for downscaling probabilistic seasonal climate forecasts - a parametric method, predictWTD, and a non-parametric method, FResampler1, and compare their performance. The predictWTD, which is based on a conditional stochastic weather generator, was found to be not very sensitive to types of rainfall information (amount, frequency or intensity) in constraining or conditioning the stochastic weather generator, but conditioning the stochastic weather generator on both rainfall frequency and rainfall intensity had distorted the distribution of the downscaled seasonal rainfall total. Both predictWTD and FResampler1 are sensitive to the length of climate data, especially for a wet SCF; climate data longer than 30 years was found suitable for reproducing the theoretical distribution of SCF. FResampler1 performed well as predictWTD in downscaling probabilistic SCF, however, it requires the generation of more realizations to ensure stable simulations of the seasonal rainfall total distributions.