Ghosal et al. (2026) Interpretable machine learning reveals contrasting drivers of streamflow droughts in regulated and natural river systems

Identification

• Journal: Ecological Indicators
• Year: 2026
• Date: 2026-09-12
• Authors: Meghomala Ghosal, Somil Swarnkar, Sudhir Kumar Singh, Vikas Poonia, Shreejit Pandey
• DOI: 10.1016/j.ecolind.2026.115501

Research Groups

• Department of Earth and Environmental Sciences, IISER Bhopal, Bhopal, Madhya Pradesh 462066, India
• K. Banerjee Centre of Atmospheric and Ocean Studies, University of Allahabad, Prayagraj, Uttar Pradesh 211002, India
• Department of Civil Engineering, MANIT Bhopal, Bhopal, Madhya Pradesh 462003, India

Short Summary

This study investigates streamflow drought dynamics in the Godavari River basin, revealing a hydro-anthropogenic dichotomy where high human-influence (HHI) sub-basins experience long-duration, moderate droughts due to regulation, while low human-influence (LHI) catchments exhibit shorter, more intense events driven by climatic forcing.

Objective

• To quantify the spatial variability of drought duration, intensity, and deficit across stations with contrasting levels of human influence.
• To develop and classify streamflow drought typologies using the Short-Duration Drought Index (SDDI) and Long-Duration Drought Index (LDDI) joint-probability indices.
• To examine the temporal evolution and relative dominance of short-duration and long-duration drought behaviour.
• To evaluate the relative contribution and nonlinear interaction of drought characteristics using Random Forest permutation importance and partial-dependence analysis to understand differences between reservoir-influenced and less-regulated catchments.

Study Configuration

• Spatial Scale: Godavari River Basin (GRB), India, covering approximately 312,812 square kilometers. The study analyzed 14 gauging stations categorized into High Human Influence (HHI) and Low Human Influence (LHI) regions.
• Temporal Scale: Daily streamflow records spanning the period 1965–2020.

Methodology and Data

• Models used:
  • Multivariate framework integrating drought duration, intensity, and deficit.
  • Joint probability approach using Empirical Cumulative Distribution Function (ECDF).
  • Inverse Generalized Extreme Value (GEV) distribution for index transformation.
  • Random Forest (RF) models for regression, permutation feature importance, and one- and two-dimensional partial dependence analyses.
  • Z-score normalization for drought characteristics.
• Data sources:
  • Open-source daily streamflow records (1965–2020) from 14 gauging stations, sourced from the India–Water Resources Information System (WRIS) and the Central Water Commission (CWC), Government of India.
  • Information on reservoirs and hydraulic structures from the National Register of Large Dams (CWC).
  • Additional geospatial datasets for basin delineation and visualization.

Main Results

• HHI sub-basins exhibit significantly longer drought durations (median exceeding 100 days) but subdued intensities, suggesting reservoir operations prolong low-flow conditions while buffering extreme drawdowns.
• LHI catchments show shorter drought durations (median generally below 80 days) but higher intensities (median ≈ 2 × 10⁻⁴ Mm³ km⁻² day⁻¹) and larger cumulative deficits, indicating greater sensitivity to climatic forcing.
• SDDI analysis reveals LHI regions are more prone to intense short-duration droughts (over 60% of events in high/very high severity classes), while HHI regions experience more low-to-moderate severity short droughts.
• LDDI analysis indicates higher frequencies of long-duration droughts in HHI regions, predominantly moderate to high severity, consistent with flow regulation extending drought persistence. LHI regions show fewer but more severe long-duration events.
• Temporal evolution (1965–2019) shows a progressive eastward shift in drought typology, with HHI zones maintaining sustained long-duration, moderate events, while LHI zones exhibit an emergence of short, severe droughts, particularly after 1990.
• Random Forest permutation importance for SDDI indicates intensity as the dominant driver in HHI regions, whereas duration plays a stronger role in LHI regions.
• For LDDI, cumulative deficit exerts the strongest control in HHI regions, while both deficit and duration contribute comparably in LHI regions.
• Two-dimensional partial dependence plots (2D PDPs) show HHI regions have smoother, saturated response surfaces, reflecting buffered drought propagation, while LHI regions exhibit sharper nonlinearities driven by unregulated climatic forcing.
• Extreme event analysis (>80th percentile) confirms regulation-induced moderation in HHI basins, flattening response surfaces, while LHI regions display strong synergistic interactions among intensity, deficit, and duration, amplifying drought severity.

Contributions

• Developed an integrated multivariate framework combining joint-probability analysis with interpretable machine learning (Random Forest) for streamflow drought characterization.
• Introduced Short-Duration Drought Index (SDDI) and Long-Duration Drought Index (LDDI) to differentiate between intensity-oriented and persistence-oriented drought regimes.
• Provided a multidimensional representation of drought severity by integrating duration, intensity, and deficit.
• Revealed contrasting drivers and typologies of streamflow droughts in regulated (HHI) versus natural (LHI) river systems within a large Indian basin.
• Demonstrated the necessity of region-specific drought monitoring and management strategies, highlighting the limitations of uniform indicators in human-influenced river systems.
• Offered a transferable framework for diagnosing drought behaviour and associated environmental risks in other large, human-influenced river systems.

Funding

• Ministry of Education Government of India (initiation grant IISERB/R&D/2022–23/186) through the Indian Institute of Science Education and Research (IISER) Bhopal.
• University Grants Commission (UGC), Government of India (fellowship support to student researcher).

Citation

@article{Ghosal2026Interpretable,
  author = {Ghosal, Meghomala and Swarnkar, Somil and Singh, Sudhir Kumar and Poonia, Vikas and Pandey, Shreejit},
  title = {Interpretable machine learning reveals contrasting drivers of streamflow droughts in regulated and natural river systems},
  journal = {Ecological Indicators},
  year = {2026},
  doi = {10.1016/j.ecolind.2026.115501},
  url = {https://doi.org/10.1016/j.ecolind.2026.115501}
}