Assessment and prediction of meteorological drought using machine learning algorithms and climate data

Abstract

Monitoring drought in semi-arid regions due to climate change is of paramount importance. This study, conducted in Morocco’s Upper Drâa Basin (UDB), analyzed data spanning from 1980 to 2019, focusing on the calculation of drought indices, specifically the Standardized Precipitation Index (SPI) and the Standardized Precipitation Evapotranspiration Index (SPEI) at multiple timescales (1, 3, 9, 12 months). Trends were assessed using statistical methods such as the Mann-Kendall test and the Sen’s Slope estimator. Four significant machine learning (ML) algorithms, including Random Forest, Voting Regressor, AdaBoost Regressor, and K-Nearest Neighbors Regressor, were evaluated to predict the SPEI values for both three and 12-month periods. The algorithms’ performance was measured using statistical indices. The study revealed that drought distribution within the UDB is not uniform, with a discernible decreasing trend in SPEI values. Notably, the four ML algorithms effectively predicted SPEI values for the specified periods. Random Forest, Voting Regressor, and AdaBoost demonstrated the highest Nash-Sutcliffe Efficiency (NSE) values, ranging from 0.74 to 0.93. In contrast, the K-Nearest Neighbors algorithm produced values within the range of 0.44 to 0.84. These research findings have the potential to provide valuable insights for water resource management experts and policymakers. However, it is imperative to enhance data collection methodologies and expand the distribution of measurement sites to improve data representativeness and reduce errors associated with local variations.

Keywords

• Drought
• Machine learning
• Upper Drâa
• Random Forest
• SPEI