Christian et al. (2025) A bi-level spatiotemporal clustering approach and its application to drought extraction

Identification

• Journal: Advances in statistical climatology, meteorology and oceanography
• Year: 2025
• Date: 2025-11-28
• Authors: Tiffany Christian, Amit N. Subrahmanya, Brandi Gamelin, Vishwas Rao, Noelle I. Samia, Julie Bessac
• DOI: 10.5194/ascmo-11-257-2025

Research Groups

• Department of Statistics and Data Science, Northwestern University, Evanston, IL, USA
• Department of Computer Science, Virginia Tech, Blacksburg, VA, USA
• Environmental Science Division, Argonne National Laboratory, Lemont, IL, USA
• Mathematics and Computer Science Division, Argonne National Laboratory, Lemont, IL, USA
• Computational Science Center, National Renewable Energy Laboratory, Golden, CO, USA

Short Summary

This paper introduces a novel bi-level spatiotemporal clustering algorithm, combining a modified space-time k-means and DBSCAN, to extract events based on their intensity and spatiotemporal structures. Applied to the Standardized Vapor Pressure Deficit Drought Index (SVDI) over the continental United States from 1980–2021, the algorithm successfully captures historical drought events and reveals long-term shifts in drought patterns.

Objective

• To develop and validate a novel flexible bi-level spatiotemporal clustering algorithm for extracting events based on their intensity and spatiotemporal structures.
• To apply this algorithm to a spatiotemporal drought index (SVDI) over the continental United States to capture historical drought events and their spatiotemporal extents.

Study Configuration

• Spatial Scale: Continental United States (longitude ∈ [−124.938°, −67.063°], latitude ∈ [25.063°, 52.938°]), discretized at 0.125° intervals (approximately 14 km). Data was downsampled by retaining every 10 grid-points for clustering. Spatial neighborhood radius of approximately 120 km (0.02 radians).
• Temporal Scale: 1 January 1980 to 31 December 2021 (42 years), with daily incremented data (29 February omitted for standardization). Temporal neighborhood window of 14 days (7 days before and 7 days after each datapoint).

Methodology and Data

• Models used:
  • Novel bi-level spatiotemporal clustering algorithm:
    • Level 1: Modified space-time k-means clustering (incorporates spatiotemporal neighborhoods into the distance metric for cluster assignment).
    • Expert-based step: Drought severity assignment (selects clusters with a mean SVDI greater than 0.5, corresponding to mild, moderate, severe, and extreme droughts).
    • Level 2: Density-Based Spatial Clustering of Applications with Noise (DBSCAN) (separates the intensity clusters into individual spatiotemporal events based on local data point densities).
• Data sources:
  • Standardized Vapor Pressure Deficit Drought Index (SVDI), calculated using daily maximum air temperature and daily minimum relative humidity.
  • North American Land Data Assimilation Systems (NLDAS) data (Xia et al., 2012) for Vapor Pressure Deficit (VPD) and subsequent SVDI calculation.

Main Results

• The proposed bi-level algorithm effectively extracts spatiotemporally coherent intensity clusters and sections them into individual events, demonstrating an improvement over standard k-means by creating temporally persistent drought regimes with realistic transitions.
• The algorithm successfully captured historically accurate drought events over the continental United States, including the 2003 central US flash drought and the 2000 southeastern flash drought. It showed robustness by identifying droughts in an unseen year (2000) with parameters tuned to a different year (2003).
• Long-term analysis (1980-2021) revealed shifts in drought patterns: mild and moderate droughts were more prevalent in the first half of the period, while severe and mild droughts predominated in the latter half.
• The frequency of mild droughts increased over time, occurring almost every year between 2000 and 2021, and their average duration also increased. Severe droughts are lasting longer and occurring more frequently than moderate droughts.
• The longest extreme drought identified in the dataset occurred during 2011.

Contributions

• Introduction of a novel flexible bi-level spatiotemporal clustering algorithm that jointly considers space and time at all steps, addressing limitations of existing methods that often treat these dimensions separately.
• Integration of a modified space-time k-means (which incorporates spatiotemporal neighborhoods) with DBSCAN to effectively extract events based on both intensity and spatiotemporal structure.
• Demonstrated ability to create temporally persistent and spatiotemporally consistent drought events, aligning with physical expectations and historical records.
• Validation of the algorithm's robustness by applying parameters tuned to one year (2003) to an unseen year (2000) with reasonable accuracy, indicating its generalizability.
• Provides a flexible and interpretable framework for partitioning large space-time environmental data into user-defined intensity classes, adaptable to various spatiotemporal scales and phenomena beyond droughts.

Funding

• Laboratory Directed Research and Development (LDRD) funding from Argonne National Laboratory, provided by the Director, Office of Science, of the U.S. Department of Energy under Contract No. DE-AC02-06CH11357.
• National Renewable Energy Laboratory, operated by Alliance for Sustainable Energy, LLC, for the U.S. Department of Energy (DOE) under Contract No. DE-AC36-08GO28308.
• DOE Office of Science Early Career Research Program.

Citation

@article{Christian2025bilevel,
  author = {Christian, Tiffany and Subrahmanya, Amit N. and Gamelin, Brandi and Rao, Vishwas and Samia, Noelle I. and Bessac, Julie},
  title = {A bi-level spatiotemporal clustering approach and its application to drought extraction},
  journal = {Advances in statistical climatology, meteorology and oceanography},
  year = {2025},
  doi = {10.5194/ascmo-11-257-2025},
  url = {https://doi.org/10.5194/ascmo-11-257-2025}
}