Detecting early warning signals of long-term water supply vulnerability using machine learning.

Adapting water resources systems to climate change requires identifying hydroclimatic signals that reliably indicate long-term transitions to vulnerable system states. While recent studies have classified the conditions under which vulnerability occurs (i.e., scenario discovery), there remains an opportunity to extend such methods into a dynamic planning context to design and assess early warning signals. This study contributes a machine learning approach to classifying the occurrence of long-term water supply vulnerability over lead times ranging from 0 to 20 years, using a case study of the northern California reservoir system. Results indicate that this approach predicts the occurrence of future vulnerabilities in validation significantly better than a random classifier, given a balanced set of training data. Accuracy decreases at longer lead times, and the most influential predictors include long-term monthly averages of reservoir storage. Dynamic early warning signals can be used to inform monitoring and detection of vulnerabilities under a changing climate. • Machine learning approaches to predict future occurrence of long-term water supply vulnerability. • Significantly outperforms benchmark random classifier over lead times ranging from 0 to 20 years. • The most influential predictors include long-term monthly averages of reservoir storage. • Dynamic early warning signals can be used to inform monitoring and detection of environmental change. [ABSTRACT FROM AUTHOR]