Deep Learning for Satellite-Based Mountain Hazard Screening: A Retrospective Failure Analysis on the 2021 Chamoli Rock-Ice Avalanche

Snigdhajyoti RakshitarXiv 2026 · arXiv

TL;DR

This paper presents a deep learning-based pipeline using U-Net and terrain analysis to screen for mountain hazards like rock-ice avalanches. By retrospectively analyzing the devastating 2021 Chamoli disaster, the authors identify critical "pipeline leaks"—specifically how terrain shadows in winter imagery can cause automated optical systems to miss high-risk detachment zones.

What Problem Does This Paper Address?

Monitoring vast, rugged mountainous regions for potential landslides or avalanches is a monumental task for human specialists. While satellite imagery offers a solution, the sheer volume of data requires automated "screening" to prioritize high-risk areas.

However, current automated models face significant hurdles:

  1. Complex Terrain: Distinguishing between rock, soil, and debris in high-altitude, snow-covered environments is difficult.
  2. Environmental Variables: Optical sensors are hampered by cloud cover and deep terrain shadows, especially during winter months.
  3. Risk Integration: Simply detecting a change isn't enough; a useful system must estimate the potential energy of a hazard and identify what infrastructure (bridges, power plants) lies in its path.

This paper evaluates a multi-stage pipeline designed to address these gaps, using the 2021 Chamoli rock-ice avalanche as a rigorous test case.

Methods

The authors developed a six-stage screening pipeline that integrates deep learning with classical geomorphological modeling:

  1. Input Preparation: A 10-band stack is created using Sentinel-2 optical bands, spectral indices (NDVI, NDSI, NDWI), and terrain slope.
  2. Surface Mapping: A U-Net architecture performs semantic segmentation, classifying pixels into six categories: background, water, snow/ice, rock, soil, and debris.
  3. Change Detection: The model compares class maps across three time steps ($t_1$ pre-event, $t_2$ post-event, and $t_3$ for persistence) to identify candidate hazard zones.
  4. Physical Quantities: For identified zones, the system calculates gravitational potential energy (GPE) using Monte Carlo sampling.
  5. Routing and Exposure: The D8 flow algorithm simulates the path a hazard would take, identifying downstream assets (from OpenStreetMap) within a 250m corridor.
  6. Ranking: Zones are ranked by a composite score of exposure, energy, and persistence to suggest a "review status" for human specialists.
Figure 2: Pipeline of screening process

Datasets

The study utilizes a diverse array of open-access data:

  • Sentinel-2 L2A: Multi-spectral imagery for surface mapping and change detection.
  • Copernicus GLO-30 DEM: Used for calculating slope, illumination, and flow routing.
  • 2m DEM Difference: High-resolution pre- and post-event elevation data (from Shean et al.) used as a "ground truth" for volume and energy calculations.
  • ERA5 Reanalysis: Climate data (precipitation and temperature) to provide environmental context.
  • OpenStreetMap (OSM): Used to build an inventory of exposed infrastructure like bridges and hydropower plants.
Table 1: Data sources and their use.

Key Insights and Ideas

  • The Shadow Problem: The most significant finding was that the optical change detection failed to identify the actual source of the Chamoli avalanche. Because the northeast-facing slope was under 73% shade prior to the incident, the U-Net could not reliably detect the surface change.
  • U-Net vs. Spectral Rules: While U-Net outperformed simple spectral thresholding in overall accuracy, it struggled specifically with the rock and soil classes in winter conditions, where spectral signatures are highly similar.
  • Ranking Logic Flaws: The current ranking system prioritized infrastructure exposure over potential energy. In the Chamoli case, this meant smaller, less dangerous zones near villages were ranked higher than the massive, high-energy detachment zone further upstream.
  • Retrospective Value: The authors argue that "retrospective failure analysis"—testing a model on a known disaster after the fact—is essential for finding "pipeline leaks" that wouldn't be apparent in standard benchmark testing.

Key Results

  • Segmentation Performance: The U-Net achieved a macro F1 score of 0.648, significantly higher than the 0.519 achieved by traditional spectral rules.
  • Class-Specific Struggles: F1 scores for rock and soil were low (around 0.3), highlighting the difficulty of distinguishing bare ground in high-altitude winter scenes.
  • Volume Capture: Using DEM differencing, the pipeline captured only 0.52 times the actual declared volume of the Chamoli event, partly due to the limitations of the 30m Copernicus DEM compared to 2m reference data.
  • Hazard Ranking: The actual source zone was ranked 14th out of 552 candidate zones within a 1km buffer, placing it in the top 2.5%, but not at the very top due to the way exposure was weighted.

Limitations and Future Work

The authors are transparent about the pipeline's current shortcomings:

  • Sunlight Dependency: The failure in shadowed areas suggests a need for SAR (Synthetic Aperture Radar) data, which can "see" through shadows and clouds.
  • Single-Event Bias: The model parameters were tuned specifically for the Chamoli scene, meaning it is not yet a general-purpose predictive tool.
  • Asset Inventory: The reliance on OpenStreetMap meant that some destroyed infrastructure was missing from the model's exposure analysis because it hadn't been mapped in the database.
  • Future Directions: The authors plan to implement multi-flow-direction routing (instead of just D8) and integrate multi-year DEM data to identify pre-failure movements in glaciers.

For more details, you can find the full code and evaluation protocols on the author's GitHub.

Summary updated

Abstract

The approach utilizes 10 bands from Sentinel-2 imagery, along with terrain information, in a deep learning pipeline to classify potential hazard sources in any given mountainous region. The performance of our model has been evaluated against the data collected during the rock-ice avalanche event in the Chamoli region in the Indian Himalayas on 7th February 2021. U-Net performs segmentation of the detected land surface in the satellite image, detects changes in the identify candidate hazard zones, and then the following steps are performed to calculate the potential energy using the Monte Carlo sampling for steepest descent (D8) flow paths, finding the exposed structures in the downstream locations, and ranking the zones depending on the risk score. In comparison to the simple spectral rule, our approach achieved the higher macro F1 score of 0.648 (as compared to 0.519). But the F1 score in case of rock and soil was low, being in the range of 0.3s. In case of optical change detection in winter, it's only that northeast-facing slope which couldn't be detected, as it was under 73% of shade prior to the incident. The parameters have been set for this particular scene alone and not independently; hence, this is a retrospective test, and this test locates a failure which has occurred earlier. The primary contribution of this paper is locating the pipeline leak.
BibTeX
@misc{rakshit2026deep,
  author = {Snigdhajyoti Rakshit},
  title = {{Deep Learning for Satellite-Based Mountain Hazard Screening: A Retrospective Failure Analysis on the 2021 Chamoli Rock-Ice Avalanche}},
  year = {2026},
  eprint = {2611.20769},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20769},
}
Figure