An autoencoder-based snow drought index.

Area of Science:

• Hydrology and Water Resources
• Climate Science
• Machine Learning Applications

Background:

• Snowmelt significantly influences global hydrology, affecting groundwater infiltration and surface runoff.
• Snow droughts, caused by diminished snowpack, pose substantial risks to water supplies, particularly in regions like the western U.S.
• Efficient detection of snow drought timing and severity is critical for water resource management.

Purpose of the Study:

• To introduce a novel indicator, the Snow Drought Response Index (SnoDRI), for identifying and quantifying snow drought events.
• To assess the efficacy of SnoDRI using historical data from the Pacific United States.
• To evaluate the importance of various snow-related variables in snow drought detection.

Main Methods:

• Developed SnoDRI using a combination of self-supervised learning (autoencoder) and mutual information.
• Employed Random Forests for feature extraction and variable importance assessment within the SnoDRI model.
• Utilized NLDAS-2 reanalysis data spanning 1981-2021 for the Pacific United States.

Main Results:

• The SnoDRI effectively identifies and quantifies snow drought occurrences.
• Variable importance analysis revealed key factors contributing to snow drought detection.
• The index's interpretation demonstrated strong coincidence with historical snow drought incidents.

Conclusions:

• SnoDRI offers a robust method for monitoring snow drought conditions.
• The developed index can enhance water resource management strategies in snow-predominant basins.
• Machine learning approaches show significant promise for advancing snow hydrology research.