Machine Learning for Attribution of Heat and Drought in Southwestern Australia
Michael B. Richman, Lance M. Leslie
Abstract
Temperature and precipitation datasets, extending back over 100 years, are analyzed at Perth, Australia. Observational analyses reveal the emergence of hot and dry years since the 1980s, with changes in maximum temperatures ~1.5-2 °C above historical means. These temperatures far exceed recorded natural variability measured in the early 20th century and, in the past few decades, have accelerated above the danger threshold established in the Paris Accords. Permutation testing of mean Perth temperature (precipitation) for the 20-year periods 1979-1998 and 1999-2018 shows an increase (decrease) of 0.855 °C (98.1 mm); p-value 0.001 (0.0087). Attribution of interannual data variability is established by wavelet analyses. Linear and support vector regression (LR, SVR), neural network (NN) and random forests (RF) are used for temperature and precipitation prediction after attribute selection methods are applied to a set of climate drivers. Forecasts on independent testing data show that for temperature and precipitation forecasts, SVR, LR and NN (temperature only) provide more accurate predictions than RF. The features selected by attribute selection and machine learning provide important guidance for climate forecasting, policy planning and management.