A Novel Approach for Downscaling Probabilistic Seasonal Climate Forecasts: Parametric or Non-Parametric?.

Because seasonal climate forecasts are commonly produced in tercile-probabilities of below-, near- and above-normal, it is inherently difficult for a practitioner to ingest and translate those quantities into more meaningful information for decision support in agriculture. In this paper, we present two new novel approaches to downscale the ‘full’ distribution of probabilistic seasonal climate forecasts and show their potential applications for managing agricultural risks associated to climate. The first approach is a non-parametric resampling and downscaling method called FResampler1 that creates weather realizations of a tercile-based seasonal climate forecasts. The method uses the concept of ‘conditional block sampling’ of weather data, based on the probabilities of rainfall forecast categories. Conditional block sampling is a way of randomly drawing, from historical records, time series of daily weather parameters e.g., rainfall, maximum and minimum temperature and solar radiation, for the season of interest from years that belong to a certain rainfall tercile category e.g., being below-, near- and above-normal. Block sampling preserves the covariance of other weather parameters with rainfall as if conditionally sampling maximum and minimum temperature and solar radiation if that day is wet or dry. The second approach is a parametric method (predictWTD) based on a conditional stochastic weather generator. The tercile-based seasonal climate forecast is converted into a theoretical forecast cumulative probability curve. The deviates for each percentile is converted into rainfall amount and/or frequency and/or intensity, and disaggregated on a monthly basis from those seasonal deviates, which are then used to constrain the downscaling of forecast realizations at different percentile values of the theoretical forecast curve. We  discuss the theoretical basis of the approach, sensitivity analysis (data volume, sample size, etc.), and case studies using actual seasonal climate forecasts for rice cropping in the Philippines, and maize cropping in India.