Spatio-temporal distribution and prediction of agricultural and meteorological drought in a Mediterranean coastal watershed via GIS and machine learning

Abstract

Drought is a complex and devastating natural disaster that needs to be constantly investigated. In this study, standardized precipitation indexes (SPI-3 and SPI-6) were computed using daily precipitation data collected from six meteorological stations in a Mediterranean coastal basin (Northwestern Morocco) during the period from 1984 to 2021. Subsequently, we examined the spatio-temporal distribution of three agricultural indices namely Vegetation Condition Index (VCI), Temperature Condition Index (TCI), and Vegetation Health Index (VHI). Additionally, to predict SPI at two scales (SPI-3 and SPI-6), we investigated the ability of four machine learning (ML) models such as random subspace, random forest, M5P, and REPTree. ML model's performance was evaluated using statistical metrics such as R2, MAE, RMSE, RAE, and RRSE. As per SPI results, 1995, 1999, 2005, 2015, and 2017 were observed as severe driest years. Agricultural drought’ magnitude differs over time and space from 1984 to 2021. Besides, findings showed that the REPTree model achieved the best performance during the testing and validation phases, with R2 (0.64–0.85), MAE (0.37–0.58), RMSE (0.51–0.74), RAE (48.10–75.34%) and RRSE (53.24–76.68%). In contrast, RF was found to offer the lowest performance accuracy, despite outperforming during the training phase. Moreover, SPI-6 has higher prediction accuracy than SPI-3. Furthermore, our findings offer a reliable model for drought prediction, which may further assist policymakers and authorities in developing better adaptation and mitigation strategies to reduce drought-related losses. In future research, we suggest exploring alternative ensemble or hybrid ML algorithms to further improve prediction accuracy and capability.