Evaluation of Flash Drought Identification with Machine Learning Techniques, Part I: Traditional Machine Learning Algorithms

Abstract

Drought is an extreme event that can have a number of impacts on water resources and agricultural productivity. To mitigate impacts, many studies have focused on improving drought predictions. As part of this, machine learning (ML) has emerged as a useful tool for drought identification and prediction, which has been shown to skillfully predict seasonal scale droughts. In addition, a subset of drought, termed flash drought, has gained increased attention due to their rapid development and difficulty to predict. Studies focused on flash droughts have found a number of variables that accurately characterize their rapidly evolving conditions. However, while much work has been conducted to investigate flash droughts, limited work has been completed using ML techniques. This study investigated the ability of ML algorithms to represent and predict flash drought events via two parts. The first part focused on standard ML techniques such as random forests, Ada boosting, and support vector machines, while the second part focused on neural network-based models. The ML models were trained on a set of key flash drought variables (soil moisture, evaporation, potential evaporation, temperature, and precipitation). Results showed Ada boosting was most effective at representing and predicting flash drought events, and was able to learn the climatological hotspot spots and some of the seasonality for flash droughts. However, the ML models tended to over rely on a few variables, and were biased towards those variables. Additionally, case studies showed the ML struggled to capture the precise timing of flash drought events, resulting in a low skill score for all the examined ML models.