DroughtSet: Understanding Drought Through Spatial-Temporal Learning

TL;DR

DroughtSet addresses the challenge of sub-seasonal to seasonal drought prediction by compiling a diverse, multi-source dataset of climate, physical, and vegetation predictors alongside three drought indices. The authors propose SPDrought, a spatial-temporal fusion model with static-dynamic feature representations and multi-task learning to forecast all three drought indices simultaneously, accompanied by an Integrated Gradient-based interpretability module. Empirical results show SPDrought outperforms strong baselines on 26-week forecasts and provide mechanistic insights into predictor importance across soil moisture, ecohydrological, and ecological droughts. The work contributes a public benchmark for climate-time-series forecasting and demonstrates how joint, interpretable predictions can enhance understanding and risk management of drought under climate change.

Abstract

Drought is one of the most destructive and expensive natural disasters, severely impacting natural resources and risks by depleting water resources and diminishing agricultural yields. Under climate change, accurately predicting drought is critical for mitigating drought-induced risks. However, the intricate interplay among the physical and biological drivers that regulate droughts limits the predictability and understanding of drought, particularly at a subseasonal to seasonal (S2S) time scale. While deep learning has been demonstrated with potential in addressing climate forecasting challenges, its application to drought prediction has received relatively less attention. In this work, we propose a new dataset, DroughtSet, which integrates relevant predictive features and three drought indices from multiple remote sensing and reanalysis datasets across the contiguous United States (CONUS). DroughtSet specifically provides the machine learning community with a new real-world dataset to benchmark drought prediction models and more generally, time-series forecasting methods. Furthermore, we propose a spatial-temporal model SPDrought to predict and interpret S2S droughts. Our model learns from the spatial and temporal information of physical and biological features to predict three types of droughts simultaneously. Multiple strategies are employed to quantify the importance of physical and biological features for drought prediction. Our results provide insights for researchers to better understand the predictability and sensitivity of drought to biological and physical conditions. We aim to contribute to the climate field by proposing a new tool to predict and understand the occurrence of droughts and provide the AI community with a new benchmark to study deep learning applications in climate science.

Figures (8)

• Figure 1: An example of drought development in July 2012. The left and right panels show the evaporative stress index in the 28th and 32nd weeks in 2012, respectively. ESI reduced in the Central Plains, indicating ecohydrological drought.
• Figure 2: SPDrought architecture for Forecasting Drought Indices: the spatial-temporal fusion module first exploits the spatial correlation of data with its neighbors using static features and leverages the learned correlation to aggregate the dynamic features; the static-dynamic feature representation exploits both spatial and temporal patterns with three network modules. Such representation is shared among multi-task regressors for generating multiple drought indices predictions. Subsequently, we analyze how individual features at various timestamps influence the final predictions using our interpretation method. Domain experts are encouraged to provide feedback on variable selection and model design, which can further refine the model and uncover deeper relationships among variables.
• Figure 3: The sensitivity of soil moisture to the top three predictive features measured by the integrated gradient, including surface pressure, radiation, and PET.
• Figure 4: The sensitivity of Evaporative Stress Index. Radiation and Pressure show a positive influence on the evaporative stress index while SIF reflects a minor negative influence.
• Figure 5: The sensitivity of Solar-induced Fluorescence. Both radiation and root-zone soil moisture directly influence the rate of photosynthesis, which in turn affects the SIF signal.