Data Bottlenecks and Algorithmic Lag in AI for Droughts: A Review of 20-Years Worth of Trends

Artificial Intelligence (AI) has emerged as a key tool in drought research, with applications growing rapidly over the past 20 years. While several reviews have described specific AI methods and their use in forecasting and monitoring, a comprehensive assessment of trends, gaps, and emerging challenges is lacking. Here, we analyze two decades of literature to map the evolution of AI in drought science. We first give a broad overview of the general trends of this field, which has grown exponentially since 2013 but has persistent gaps in drought-prone regions such as Africa and South America. We then discuss trends and biases in the algorithms and data used. Specifically, we find that the field has not yet fully leveraged the “big data” revolution, and instead relies mostly on small datasets focused on local case studies. Furthermore, there seems to be no upward trend in the average size of the datasets used in studies on AI for droughts. This lack of large-scale benchmarks has two profound consequences. First, it makes it challenging to compare the performance of algorithms trained on different case studies, as their generalizability cannot be assumed a priori. Second, the lack of large datasets has inadvertently constrained the field to relying, for the most part, on simple machine learning algorithms, making it lag behind in the adoption of more advanced, state-of-the-art AI tools. For example, while the general AI literature saw an exponential increase in the use of convolutional neural networks after ~2012, this type of algorithm first appeared in the literature on AI for droughts in 2019, and has seen timid uptake since then. Finally, we discuss reproducibility challenges due to restricted code and data sharing. Addressing these issues is critical to advance AI-based drought risk management and climate adaptation.