Drought forecasting using a hybrid neural architecture for integrating time series and static data
Julian Agudelo, Vincent Guigue, Cristina Manfredotti, Hadrien Piot
TL;DR
The paper addresses drought forecasting across heterogeneous regions by proposing a hybrid neural architecture that fuses time-series meteorological data with static county-level descriptors to predict USDM drought categories. The model combines LSTM-based time-series modeling, categorical embeddings, a feedforward path for static features, and an attention mechanism, producing six-week forecasts $\hat{y} \in \mathbb{R}^6$. It demonstrates state-of-the-art performance on the DroughtED dataset, supported by ablation analyses showing the importance of attention and meteorological data, and reveals that location-agnostic training offers robust generalization across counties. The work highlights a modular framework capable of leveraging heterogeneous data for climate-related tasks and points to practical advances for early warning and adaptive drought management, with future work on attention calibration and expert-supervised attention alignment.
Abstract
Reliable forecasting is critical for early warning systems and adaptive drought management. Most previous deep learning approaches focus solely on homogeneous regions and rely on single-structured data. This paper presents a hybrid neural architecture that integrates time series and static data, achieving state-of-the-art performance on the DroughtED dataset. Our results illustrate the potential of designing neural models for the treatment of heterogeneous data in climate related tasks and present reliable prediction of USDM categories, an expert-informed drought metric. Furthermore, this work validates the potential of DroughtED for enabling location-agnostic training of deep learning models.