Multiscale determination and characterization of drought precursors in Córdoba province (Southern Spain) using machine learning models

The geographical area covered by this work is Córdoba province, located in Southern Spain. This region usually faces an extremely severe drought situation, exacerbated by the current climate emergency that predicts more frequent and stronger events. Historical droughts, including flash droughts, could be preceded by patterns such as heat waves, alternating periods of flooding and drought, peak vapor pressure deficits (VPDs), as well as incipient decline of the vegetation condition and soil moisture content, among other indicators. This study aims to conduct a retrospective multiscale analysis to identify and characterize these precursors using machine learning methods, relating the main drought indices from the scientific literature to various atmospheric and agrometeorological indicators from previous time periods. The methodology includes: 1) The use of high-quality input data, 2) Feature selection, 3) Model selection (Multilayer Perceptron, Extreme Learning Machine, Random Forest and Support Vector Machine), 4) Hyperparameter tuning, 5) Evaluation and validation. To ensure generalization capability of the models forecasting successive new datasets, robust evaluation is carried out. The model performances will be assessed by using the statistical parameters commonly used in studies of this type and considering the most current recommendations. The results obtained could be used to improve the predictive ability for droughts and, with them, the appositeness of the present early warning systems. Similarly, the methodology used in this study could be replicated in other vulnerable areas of interest, regardless of their geographical location.