On the use of convolutional Gaussian processes to improve the seasonal forecasting of precipitation and temperature

This study examines the potential improvement in seasonal predictability of monthly precipitation and temperature using a novel machine learning approach, the convolutional Gaussian process (CGP). This approach allows us to take into account multiple quantities and their interdependencies simultaneously. We use one global climate model (FLORb01) part of the North American Multi-Model Ensemble (NMME) project and quantify its skill in reproducing precipitation and temperature in March and July across Iowa (central United States) for lead times from one month to one year. As a first step we train the CGP over the 1985–2005 period, and then apply it out of sample from 2006 to 2019. Over the validation period, our results indicate that the CGP is able to increase the skill (i.e., increased correlation coefficient and reduced root mean squared error) in predicting precipitation and temperature compared to both the raw outputs and after standard bias correction. These statements are consistent across different lead times and target month (i.e., March or July). These encouraging findings provide a new potential path towards improved predictability of the regional climate at the seasonal scale.