Machine learning algorithms for the prediction of drought conditions in the Wami River sub-catchment, Tanzania

Abstract

Study region: This study refers to the Wami river sub-catchments in Eastern Tanzania. Study Focus: The five-machine learning (ML) algorithms, including long short-term memory (LSTM), multivariate adaptive regression spline (MARS), support vector machine (SVM), extreme learning machine (ELM), and M5 Tree, were used to predict the most widely used drought index, the standard precipitation index (SPI), at six and nine months. Algorithms were established using monthly rainfall data for the period from 1990 to 2022 at five meteorological stations distributed across the Wami River sub-catchment: Barega, Dakawa, Dodoma, Kongwa, and Mandera stations. New hydrological insights for the region. The predicted results of all five ML algorithms were evaluated using several statistical metrics, including Pearson's correlation coefficient (R), mean absolute error (MAE), root mean square error (RMSE), and Nash Sutcliffe efficiency (NSE). The prediction results revealed that LSTM perform better in predicting drought conditions using SPI6 (6-month SPI) and SPI9 (9-month SPI) with the highest NSE of 0.99 in all five stations, and R of 0.99 in four stations except at Kongwa station, where R range from 0.75 to 0.99. These prediction results will aid decision-makers and planners to develop a drought monitoring and drought early warning system in order to strengthen the governance and resilience to the catchment and people on the impacts of water scarcity and climate change.