Forecasting drought using machine learning: a systematic literature review

• Autor(a) principal: Oyarzabal, Ricardo S.
• Data de Publicação: 2025
• Outros Autores: Santos, Leonardo B. L., Cunningham, Christopher, Broedel, Elisangela, de Lima, Glauston R. T., Cunha-Zeri, Gisleine, Peixoto, Jerusa S., Anochi, Juliana A., Garcia, Klaifer, Costa, Lidiane C. O., Pampuch, Luana A. [UNESP], Cuartas, Luz Adriana, Zeri, Marcelo, Guedes, Marcia R. G., Negri, Rogério G. [UNESP], Muñoz, Viviana A., Cunha, Ana Paula M. A.
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1007/s11069-025-07195-2; https://hdl.handle.net/11449/300634
• topic: Climate change; Drought forecasting; Machine learning; Systematic literature review

The number of reported drought events per year and their impacts have significantly increased in the last two decades. In addition to monitoring drought conditions, forecasting is essential for planning activities. Various Machine Learning (ML) algorithms have experienced a substantial increase in popularity in geoscience applications. This study presents a Systematic Literature Review on drought forecasting utilizing Machine Learning models. Following the PRISMA 2020 protocol (Preferred Reporting Items for Systematic Reviews and Meta-Analyses), the total number of papers was reduced from approximately a thousand to a hundred. The majority of the papers found study areas from Asia and Oceania. Meteorological drought was the most studied event in the articles evaluated due to the greater ease of its estimation using only rainfall data. The Standardized Precipitation Index and the Standardized Precipitation Evapotranspiration Index are the most widely used indices in research relating to drought and Machine Learning. Precipitation is the most commonly used input among the various input data used in ML models. Remote sensing has yet to be widely used in drought forecasting, with less than 20% of papers utilizing remote sensing data. What still needs to be addressed is drought forecasting in the time scale of days, which is less utilized compared to the monthly scale. The regression method is the most commonly used, with 77% of papers utilizing it. In conclusion, we formulated five recommendations based on the critical evidence and insights from our review: (1) it is essential to foster interdisciplinary collaborations among experts in ML, climatology, and hydrology while investing in initiatives that promote the sharing of data and code repositories; (2) satellite remote sensing technologies and crowd-sourced data collection methods should be considered in ML studies while enhancing existing monitoring infrastructure to increase the spatial and temporal coverage of datasets for validation of ML methods; (3) it is recommended to increase the availability of additional environmental variables, such as soil moisture and vegetation health, to promote more studies of agricultural drought and ML methods; (4) it is crucial to prioritize the integration of daily-scale climate data into drought modeling and forecasting for developing effective adaptation and mitigation measures to flash drought events; and finally (5) ethical considerations of using Artificial Intelligence (AI) for drought forecasting, emphasizing the environmental impact, issues of digital sovereignty, and the urgent need for a broader dialogue on AI’s role in sustainable climate solutions.