Spatio-Temporal Modelling for Downscaling Long-Range Daily Precipitation Forecasts at Unmonitored Sites

Downscaling low-resolution long-range climate forecasts to specific sites, in the absence of historical observations, poses a challenge yet holds significant value for weather-sensitive sectors. Existing methods for statistical downscaling and climatological forecasts, however, necessitate site-specific observations. In this study, we propose a novel solution for generating daily precipitation forecasts at any unmonitored site. We introduce a Bayesian non-stationary truncated spatio-temporal model that uses data from meteorological stations within its surrounding region. The model employs a truncated random process to handle dry days and utilises a Gaussian-process-based spatio-temporal stochastic process to capture non-stationary spatial dependencies. Besides utilising ensemble seasonal forecasts and hindcasts from a climate model, it also incorporates Quantile Mapping (QM) downscaled long-range forecasts for these stations. Our proposed model, Spatio-Temporal Forecast downscaling with QM (STFQM), is employed to downscale climate forecasts from ACCESS-S1 (Australian Community Climate and Earth-System Simulator – Seasonal, version 1) at a 60km spatial resolution to any site within the Murray-Darling basin, Australia. Through comprehensive leave-one-station-out cross-validation, our model demonstrates superior performance in one-month lead time forecasts compared to both the raw ACCESS-S1 forecasts and their QM downscaled counterparts, in terms of accuracy, reliability, and overall ensemble forecast skill. Furthermore, our results illustrate that STFQM is on par with Climatology, a benchmark for seasonal forecasts. Our innovative approach complements existing statistical downscaling techniques, eliminating the need for site-specific observations while providing high-quality localised precipitation forecasts.