No abstract

Drought is a costly natural disaster that impacts economies and ecosystems worldwide, so monitoring drought and communicating its impacts to individuals, communities, industry, and governments is important for mitigation, adaptation, and decision‐making. This research describes a novel machine learning framework to predict and understand the Canadian Drought Monitor (CDM). This fully automated approach is trained on nearly two decades of expert analysis and would assist the comprehensive monitoring of drought impacts without the continued requirement of ground support, a benefit in many data‐limited areas across the country. The framework also integrates the Shapley Additive Explanation (SHAP) variable importance metric to provide insight into drought dynamics in near real‐time, demonstrating its usefulness for understanding the value of different data sets for drought assessments and dispelling the commonly held misconception that machine learning models are not useful for inference. The results demonstrate that the model can effectively predict the CDM maps and realistically capture the evolution of drought events over time. A SHAP analysis found that the Prairie drought of 2015 was related to a strong El Niño event that reduced water supply to a region already facing long‐term water deficits, and the subsequent reduction in groundwater availability was detected by the Gravity Recovery and Climate Experiment satellite. Overall, this research shows strong potential to streamline the CDM methodology, integrate scientific insight into operations in near real‐time using SHAP values, and provide an avenue to retrospectively extend the CDM for evaluating current and future drought events in a historical context.

Drought is a costly natural disaster characterized by water shortages that impact water availability, agriculture, ecosystems, and the economy. The driving mechanisms of drought operate on a wide range of spatial scales, from the movement of soil water on a hillslope to global atmospheric circulation. Additionally, drought impacts vary across spatial scales, from drought induced crop stress on a specific agricultural field to widespread continental water shortages. As a result, multi-scalar drought monitoring and early warning systems are needed to utilize observational datasets obtained at different spatial scales and to communicate drought impacts to various levels of decision-makers in government and industry. However, scaling must be employed to translate information across scales, either to fix incongruencies in the spatial scale of input datasets or to modify the model output scale. These scaling techniques have several challenges and limitations that hinder drought accuracy and interpretability, such as the Modifiable Areal Unit Problem (MAUP) and increased model uncertainty. This paper reviews the role of spatial scale in drought monitoring and early warning systems, the associated challenges, and techniques to minimize their impact. Finally, this review identifies several knowledge gaps and future directions.