Interpretable Deep Learning Reveals Influence and Interactions of Hydroclimatic Drivers on Floods Across India
Earth's Future · 2026-07 · 14 (7)
DOI: 10.1029/2026ef008397 ↗Kurz gefasst
Fachartikel in der Zeitschrift Earth's Future, online erschienen am 22.07.2026. Verfasst von Vaibhav Tripathi und 4 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Department of Water Resources, Indian Institute of Technology Roorkee.
Abstract (Originalsprache)
River flooding emerges from complex interactions among rainfall intensity, antecedent wetness, and human interventions, yet the spatial organization of these processes that leads to extreme floods remains poorly understood, particularly in physiographically diverse regions such as India. Here, we present a first nation‐wide, event‐based classification of flood‐generating processes and develop an interpretable deep‐learning framework to predict extreme river‐flood magnitudes across 206 catchments in peninsular India. Flood events during 1980–2020 are classified into three hydro‐meteorological process types: long‐rain, excess‐rain, and short‐rain floods, based on joint rainfall–soil‐moisture thresholds. We investigate how the dominance of these processes varies with hydroclimate and reservoir storage. Cluster‐specific TabNet models are trained to predict flood magnitude from daily hydro‐meteorological forcings and basin‐averaged static attributes under strict spatial cross‐validation. SHAP values are derived to attribute predictions to dominant drivers and their interactions. Results reveal the dominant flood‐generating process systematically shifts with climate and reservoir storage: humid, limited‐storage basins are controlled by excess rain, whereas drier storage‐influenced catchments show a much larger contribution of long‐rain floods, revealing a joint control of aridity and catchment storage on how rainfall is transformed into floods. The cluster‐wise models achieve moderate skill, with ∼67% of basins achieving KGE >0.3 and reveal physically consistent driver combinations: intensity plus surface wetness dominate excess‐rain regimes, while multi‐day rainfall and deep soil moisture dominate long‐rain regimes. These findings advance understanding of flood processes and demonstrate how interpretable deep learning enhances hydrologic prediction, establishing a framework for climate‐adaptive flood forecasting in monsoon regions worldwide.
Förderung
- Department of Water Resources
- Indian Institute of Technology Roorkee
Zitieren
APA
Tripathi, V., AghaKouchak, A., Karmakar, S., Stein, L., & Mohanty, M. P. (2026). Interpretable Deep Learning Reveals Influence and Interactions of Hydroclimatic Drivers on Floods Across India. Earth's Future, 14(7). https://doi.org/10.1029/2026ef008397
Vancouver
Tripathi V, AghaKouchak A, Karmakar S, Stein L, Mohanty MP. Interpretable Deep Learning Reveals Influence and Interactions of Hydroclimatic Drivers on Floods Across India. Earth's Future. 2026;14(7). doi:10.1029/2026ef008397
BibTeX
@article{Tripathi2026,
author = {Tripathi, Vaibhav and AghaKouchak, Amir and Karmakar, Subhankar and Stein, Lina and Mohanty, Mohit Prakash},
title = {Interpretable Deep Learning Reveals Influence and Interactions of Hydroclimatic Drivers on Floods Across India},
journal = {Earth's Future},
year = {2026},
volume = {14},
number = {7},
doi = {10.1029/2026ef008397}
}
EUVIMED