Probabilistic Evaluation of Drought Propagation Using Satellite Data and Deep Learning Model: From Precipitation to Soil Moisture and Groundwater

Abstract

The frequency of drought events has increased with climate change, making it vital to monitor and predict the response to drought. In particular, the relationship among meteorological, agricultural, and groundwater droughts needs to be characterized under different drought conditions. In this study, a probabilistic framework was developed for analyzing the spatio-temporal propagation of droughts and applied to South Korea. Three drought indices were calculated using satellite data and a deep learning model to determine the spatial and temporal extents of drought. The average propagation times were calculated. The time from meteorological to agricultural drought (MD-to-AD) was 2.83 months, and that from meteorological to groundwater drought (MD-to-GD) was 4.34 months. Next, the joint distribution among three drought types based on the best-fit copula functions was constructed. The conditional probabilities of drought occurrence were calculated on temporal and spatial scales. For instance, the probabilities of MD-to-GD propagation under light, moderate, severe, and extreme meteorological drought conditions were 38%, 43%, 48%, and 53%, respectively. The propagated drought occurrence probability was confirmed to be the highest under extreme antecedent drought conditions. The results of this study provide insight into the spatio-temporal drought propagation process from a probabilistic viewpoint. The use of satellite data and a deep learning model is expected to increase the efficiency of drought management practices such as vulnerability assessment and early warning system development.

Keywords

• Deep learning
• groundwater drought
• probability
• propagation
• satellite