Regional Drought Index based on Machine Learning Models

Abstract

Global warming has intensified the frequency and severity of droughts worldwide, complicating the characterization and continuous monitoring of regional drought conditions. Traditional methods often struggle with the spatial variability of meteorological factors and the influence of extreme weather events, leading to potential distortions in drought analysis. This thesis introduces two innovative drought indices aimed at enhancing the accuracy and reliability of regional drought monitoring: the Support Vector Machine-based Drought Index (SVM-DI) and the Auxiliary Support Vector Machine Drought Index (Aux-SVMDI). The SVM-DI is developed by applying different weights to an SVM-based X-bar control chart, incorporating regional precipitation aggregate data. This new index, tested in northern Pakistan, demonstrates more pronounced regional characteristics and a significantly lower Coefficient of Variation (CV) in its correlations with other meteorological stations compared to the Regional Standard Precipitation Index (RSPI). These findings highlight SVM-DI's capability to reduce the impact of extreme values and outliers, making it a robust tool for regional drought analysis. Similarly, the Aux-SVMDI integrates monthly precipitation estimates derived from auxiliary data and predicted errors using SVM, combined with temperature as an auxiliary variable in a regression estimation framework. This index employs control charts derived from SVM error estimations to enhance sensitivity in detecting hydrological droughts. The results indicate that Aux-SVMDI provides superior estimates by utilizing auxiliary information, thereby offering a more precise and reliable assessment of hydrological drought conditions. Both SVM-DI and Aux-SVMDI represent significant advancements in drought monitoring methodologies, offering enhanced tools for practitioners to accurately define and manage regional climatology and drought conditions. Future research should explore the integration of additional climatic variables and advanced machine learning techniques to further refine these indices.