Machine learning and remote sensing based time series analysis for drought risk prediction in Borena Zone, Southwest Ethiopia

Abstract

Drought, a complex natural hazard, poses significant challenges for scholars in drought risk management due to the perceived lack of suitable assessment, prediction, and monitoring tools. Addressing these challenges requires sophisticated tools capable of precise and timely assessments. This study leverages the rapid advancements in machine learning (ML) and remote sensing (RS) to develop models for anticipating and estimating drought risk, particularly focusing on identifying affected and sensitive locations in Ethiopia's Borena Zone. The research involves a two-phase investigation, examining historical drought patterns and forecasting scenarios for 2028. The GridSearch algorithm is employed for optimal hyperparameter tuning in ML, highlighting the CatBoost algorithm as the most accurate predictor for the Standardized Precipitation Index (SPI). With impressive performance metrics, including Mean Squared Error (MSE) of 0.017, Mean Absolute Error (MAE) of 0.102, Root Mean Square Error (RMSE) of 0.129, and an R-squared (R²) value of 0.84, this study excels in providing precise spatiotemporal accuracy for drought prediction. The findings underscore the importance of time-series drought prediction, offering crucial insights for decision-makers and planners to address and mitigate drought impacts at various scales. This study contributes valuable information by emphasizing the significance of understanding drought occurrence's temporal and spatial dimensions.