An impact-oriented framework for a deep learning–based composite drought index considering potential economic losses

Bibliographic Data

ID15544832
AuthorsMostafa Khosh Chehreh (0000-0001-7667-507X, Politecnico di Milano, corresponding author), Carlo De Michele (0000-0002-7098-4725, Politecnico di Milano)
Year2025
Volume20
Issue12
Pages124069-124069
Publication date2025-12-01
Peer ReviewedYes
Open AccessYes
TypeARTICLE
VenueEnvironmental Research Letters (JOURNAL)
Journal identifiersISSN: 1748-9326 • E-ISSN: 1748-9326
PublisherIOP Publishing (PUBLISHER • GB)
DOI10.1088/1748-9326/ae269a
OpenAlexW4416929399
LanguageEN
References cited65

Droughts significantly impact socioeconomic systems globally. As a complex and multifaceted phenomenon, drought is characterized through various indices—meteorological, agricultural, and hydrological- each capturing different aspects of the phenomenon. This diversity has led to growing demand for integrated drought monitoring tools that offer a more holistic understanding of drought conditions. Traditionally, newly developed composite drought indices are assessed through the comparison with existing indices. However, this model-by-model validation approach does not necessarily reflect real-world accuracy or relevance. Therefore, a paradigm shift is needed—from comparative validation among indices to impact-oriented evaluation—emphasizing the capacity of drought indices to capture actual societal and environmental consequences. In this study, we propose a novel drought index derived from deep learning, evaluated through an impact-oriented lens using drought-induced economic losses as the primary performance metric. The index is computed using advanced deep learning techniques and a range of drought-related variables. To enhance model accuracy and robustness, we employ different self-supervised learning architectures, including convolutional neural networks, artificial neural networks, and variational autoencoders. The analysis utilizes ERA5 reanalysis data (1989–2024) for Italy, integrated with economic loss records from the EM-DAT database. Each model’s performance is assessed based on its ability to estimate potential economic losses caused by drought. The proposed framework allows users to select the most suitable index based on a guided analysis, balancing data collection effort and predictive reliability

Artificial neural network · Convolutional neural network · Data collection · Deep learning · Index (typography · Range (aeronautics · Hydrology and Drought Analysis · Logistics and Infrastructure Analysis · Remote Sensing in Agriculture

Tools

Open DOIOpen Access