Checking the Reliability of Deep Learning Climate Predictive Model

It is well known that checking the reliability of a statistical climate predictive model is necessary. This study investigates the reliability of a deep learning (DL) climate predictive model. The reliability is investigated by checking the interpretability of the model and the stability of the predictive skill. The results of multi group ensemble prediction show that the predictive skill of training data and that of testing data have large disturbance, and the disturbance amplitude increases with increasing forecast lead months. The sensitivity of predictands to predictors in the DL model is inconsistent with the background knowledge and even absurd. It is clear that the DL model is not interpretable. In short, the case analysis suggests that the DL climate predictive model might be not reasonable even if the model can produce higher predictive skill than the physical model. Like the traditional linear statistical predictive model, checking the reliability of the DL predictive model is also very necessary.