Development of a Composite Drought Index using deep learning: A Unified Framework for Multi-Dimensional Drought Characterization

 

Abstract. Droughts significantly affect socioeconomic conditions globally. As a multifaceted phenomenon, droughts are assessed through various indices distinguished in meteorological, agricultural, and hydrological typologies, each designed to capture distinct aspects. So, there is a strong demand for comprehensive drought monitoring tools that integrate multiple aspects to offer a holistic view of drought conditions. Typically, when introducing a new composite drought index, it is evaluated in comparison to existing indices, however this approach cannot allow to evaluate its accuracy in actual conditions. Therefore, shifting the paradigm from model-by-model evaluations to impact-oriented analysis is crucial. This work introduces a drought index based on deep learning where economic losses induced by drought are used as a key metric in assessing the index performance. The introduced index is calculated using cutting-edge deep learning algorithms based on various drought-related variables. Different types of self-supervised learning models, including Convolutional Neural Networks (CNN), Artificial Neural Networks (ANN), and Variational Autoencoders, are employed to enhance the model's accuracy and robustness. We use reanalysis data (ERA5) spanning from 1980 to 2022 for Italy, coupled with the EM-DAT database, to conduct impact analysis. The performance of each model is outlined based on their accuracy in estimating economic losses induced by droughts. 

Keywords: Drought Index, Deep learning, Autoencoder, impact-oriented analysis.