Drought risk assessment: integrating meteorological, hydrological, agricultural and socio-economic factors using ensemble models and geospatial techniques

Among natural disasters, drought hits almost half of the world every year, regardless of the climatic zones. Identifying drought vulnerability regions is fundamental to plan and adopt mitigation measures. Here we apply a multi-criteria-based machine learning technique that integrates spatial data for preparing drought vulnerability map of different categories. We adopted remote sensing tools with three machine learning models namely support vector machine (SVM), random forest (RF) and support vector regression (SVR) and their ensembles (i.e. Bagging, Boosting and Stacking), as applied to the northwestern part of Iran as a case study. Various types of geo-environmental factors were considered including meteorological, hydrological, agricultural and socio-economic. The result of the model was evaluated through arithmetic logic values (area under the curve [AUC]) under the receiver operating curve (ROC). Through multi-collinearity test, the prominent causative factors for the occurrences of drought are defined. The AUC value from ROC of SVR-Stacking, RF-Stacking and SVM-Stacking model for training datasets are 0.942, 0.918 and 0.896, respectively. The SVR-Stacking yielded the best result (AUC = 0.94) confirming that SVR serves as a robust model for the preparation of drought susceptibility maps that can be used by governmental and other administrative agencies.

Automated summary

This study utilized a machine learning technique to create drought vulnerability maps in northwestern Iran, with the SVR model proving to be the most effective in predicting drought susceptibility. The results highlight the importance of identifying causative factors for drought occurrences to aid in mitigation efforts.