Latitudinal variations in the accuracy of model-generated forecasts of precipitation over Australia and south-east Asia

Abstract

Forecasts of precipitation produced by global and regional versions of the Bureau of Meteorology’s Australian Community Climate and Earth System Simulator (ACCESS) numerical weather prediction models are compared with Tropical Rainfall Measuring Mission (TRMM) observations for the period January 2011 to March 2014. The area considered covers longitudes 110° to 160°E within 40° latitude of the equator, and includes the Australian continent and part of south-east Asia. Forecast accuracy is assessed using objective measures: equitable threat score (ETS), frequency bias (FB), and the ratio of predicted to observed rain volume (RVR). For the assessment, the TRMM and model datasets are both interpolated to consistent grids defined by spacings of 1° longitude and 1° latitude. Assessments based on three-month seasons show that, in general, the volume of rain predicted by the models is too high, and rainfall is predicted to occur at more locations than observed. Latitudinal variations in values of the ETS reveal marked declines over the equatorial tropics, with minimum values near the equator from December to May and near 10°N from June to November. Differences in the ETS between 24-hour model and persistence forecasts show that the models produce useful predictions of rainfall away from the tropics, poleward of latitudes 5º-15º S and 15º-30º N, depending on the season and model. The 48-hour model predictions are more useful than the 24-hour forecasts, with improvements relative to persistence over most latitudes, although differences are close to zero at low latitudes. Varying the rain threshold used to compute skill metrics shows that the models produce excessive rainfall at low rain rates, generally less than approximately 15 mm day^-1, while not enough precipitation is forecast at higher rain rates. Values of the ETS are generally highest at lower rain rates, coinciding with the excessive values of RVR and FB.