Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Flood Risk Follows Valleys, Not Grids: Graph Neural Networks for Flash Flood Susceptibility Mapping in Himachal Pradesh with Conformal Uncertainty Quantification

Paras Sharma, Swastika Sharma

Abstract

Flash floods are the most destructive natural hazard in Himachal Pradesh (HP), India, causing over 400 fatalities and $1.2 billion in losses in the 2023 monsoon season alone. Existing risk maps treat every pixel independently, ignoring the basic fact that flooding upstream raises risk downstream. We address this with a Graph Neural Network (GraphSAGE) trained on a watershed connectivity graph (460 sub-watersheds, 1,700 directed edges), built from a six-year Sentinel-1 SAR flood inventory (2018-2023, 3,000 events) and 12 environmental variables at 30 m resolution. Four pixel-based ML models (RF, XGBoost, LightGBM, stacking ensemble) serve as baselines. All models are evaluated with leave-one-basin-out spatial cross-validation to avoid the 5-15% AUC inflation of random splits. Conformal prediction produces the first HP susceptibility maps with statistically guaranteed 90% coverage intervals. The GNN achieved AUC = 0.978 +/- 0.017, outperforming the best baseline (AUC = 0.881) and the published HP benchmark (AUC = 0.88). The +0.097 gain confirms that river connectivity carries predictive signal that pixel-based models miss. High-susceptibility zones overlap 1,457 km of highways (including 217 km of the Manali-Leh corridor), 2,759 bridges, and 4 major hydroelectric installations. Conformal intervals achieved 82.9% empirical coverage on the held-out 2023 test set; lower coverage in high-risk zones (45-59%) points to SAR label noise as a target for future work.

Flood Risk Follows Valleys, Not Grids: Graph Neural Networks for Flash Flood Susceptibility Mapping in Himachal Pradesh with Conformal Uncertainty Quantification

Abstract

Flash floods are the most destructive natural hazard in Himachal Pradesh (HP), India, causing over 400 fatalities and $1.2 billion in losses in the 2023 monsoon season alone. Existing risk maps treat every pixel independently, ignoring the basic fact that flooding upstream raises risk downstream. We address this with a Graph Neural Network (GraphSAGE) trained on a watershed connectivity graph (460 sub-watersheds, 1,700 directed edges), built from a six-year Sentinel-1 SAR flood inventory (2018-2023, 3,000 events) and 12 environmental variables at 30 m resolution. Four pixel-based ML models (RF, XGBoost, LightGBM, stacking ensemble) serve as baselines. All models are evaluated with leave-one-basin-out spatial cross-validation to avoid the 5-15% AUC inflation of random splits. Conformal prediction produces the first HP susceptibility maps with statistically guaranteed 90% coverage intervals. The GNN achieved AUC = 0.978 +/- 0.017, outperforming the best baseline (AUC = 0.881) and the published HP benchmark (AUC = 0.88). The +0.097 gain confirms that river connectivity carries predictive signal that pixel-based models miss. High-susceptibility zones overlap 1,457 km of highways (including 217 km of the Manali-Leh corridor), 2,759 bridges, and 4 major hydroelectric installations. Conformal intervals achieved 82.9% empirical coverage on the held-out 2023 test set; lower coverage in high-risk zones (45-59%) points to SAR label noise as a target for future work.
Paper Structure (52 sections, 4 equations, 12 figures, 3 tables)

This paper contains 52 sections, 4 equations, 12 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Research Gaps
  3. Gap 1 --- No comprehensive ML study for HP.
  4. Gap 2 --- Spatial autocorrelation in validation.
  5. Gap 3 --- No uncertainty quantification.
  6. Gap 4 --- Absence of graph-based spatial dependencies.
  7. Gap 5 --- Cloud-contaminated or absent flood inventories.
  8. Objectives and Contributions
  9. Study Area
  10. Geographic Setting
  11. Physiographic Zones
  12. Hydroclimatology
  13. Flood History and Significance
  14. Data and Methods
  15. Flash Flood Inventory
  16. ...and 37 more sections

Figures (12)

  • Figure 1: Study area: Himachal Pradesh, India. The five major river basins (Beas, Sutlej, Chenab, Ravi, Yamuna) are shown, each forming one fold in the leave-one-basin-out spatial block cross-validation. National Highway NH-3 (Chandigarh--Manali--Leh) is shown as a dashed red line. Inset: location of HP within India.
  • Figure 2: Methodological workflow. Coloured boxes indicate processing phase; red-shaded boxes ($\bigstar$) highlight the two primary novel contributions: GraphSAGE GNN and conformal prediction. Small-patch morphological filtering is performed post-download in Python using scipy binary opening.
  • Figure 3: ROC curves for baseline pixel-based models under leave-one-basin-out spatial block cross-validation. Each curve represents one spatial block fold. The stacking ensemble achieves the best pixel-level AUC (0.881), with notable variation across folds; the Trans-Himalayan block (basin_0) is consistently the most challenging.
  • Figure 4: Model comparison under leave-one-basin-out spatial block CV. Box plots of AUC-ROC across five basin folds for all models (left) and mean AUC $\pm$ SD per model (right). Dashed line: saha2023 benchmark (AUC = 0.88, Beas basin, random split). GNN-GraphSAGE outperforms all pixel-based baselines in all folds.
  • Figure 5: AUC improvement from watershed graph structure. Grouped bars show GNN-GraphSAGE vs. best pixel-based model (Stacking Ensemble) per k-means spatial block fold. The GNN improves over the pixel-based baseline in all five folds. The largest absolute gain (+0.27) occurs in basin_0, a predominantly Trans-Himalayan block where pixel-based models struggle most (AUC $\approx$ 0.73).
  • ...and 7 more figures