Abstract
Spatially dense and continuous information on avalanche occurrences is crucial for numerous safety-related applications such as avalanche warning, hazard zoning, hazard mitigation measures, forestry, risk management, and numerical simulations. Currently, this information is collected non-systematically by field observers. While remote sensing technology offers a way to fill this gap, the manual mapping process remains a significant bottleneck. This study deploys a modified DeepLabV3+ deep learning model to automatically identify and map avalanches in SPOT 6/7 imagery. Using 24,778 manually annotated polygons for training and validation, the model achieved an average F1-score of 0.625, comparable to the reproducibility of human experts. This approach represents a significant step toward rapid, comprehensive documentation of avalanche periods, enhancing safety in mountain regions.
Introduction
Accurate records of avalanche timing and location are vital for safety warnings, hazard zoning, and infrastructure planning. Traditionally, these data are collected unsystematically by field observers, leading to gaps in spatial and temporal coverage. Satellite remote sensing, specifically high-resolution optical imagery, provides a means to monitor large regions simultaneously. However, manual digitization of avalanche outlines is labor-intensive; an expert typically requires one hour to map just 27.5 km². To monitor entire mountain ranges like the Alps effectively after major storm events, automated solutions are required.
Methodology
The researchers employed DeepLabV3+, a state-of-the-art semantic segmentation model, enhanced with two "terrain-aware" modifications designed to account for the physical nature of avalanches:
- Deformable Convolutions: Unlike standard rigid kernels, these adapt their shape based on the Digital Elevation Model (DEM). This allows the network to effectively "feel" the slope and terrain features that guide avalanche flow.
- Deformable Spatial Pyramid Flow (DSPF): This module helps the model propagate information along likely downhill flow directions, mimicking the physical movement of snow.
The model was trained using a weighted binary cross-entropy loss, prioritizing pixels from "well-visible" avalanches to improve the quality of the learned features.
Data and Study Area
The study utilized a robust dataset comprising:
- SPOT 6/7 Satellite Imagery: Optical data with a 1.5m ground sampling distance (RGB and Near-Infrared bands).
- swissALTI3D: A high-resolution Digital Elevation Model of Switzerland.
- Manual Annotations: A massive dataset of 24,778 manually annotated avalanche polygons from major events in January 2018 and January 2019.
- Generalization Set: Imagery from January 6, 2018, was reserved to test the model's performance on unseen temporal data.
Results and Performance
The model demonstrated strong performance across several metrics:
- Pixel-level Metrics: The model achieved an average Probability of Detection (POD) of 0.610, a Positive Predictive Value (PPV) of 0.668, and an overall F1-score of 0.625.
- Object-level Detection: Approximately 66% of individual avalanches were correctly identified (using a 50% overlap threshold).
- Generalization: Performance remained stable when applied to a completely different avalanche period, indicating the model's robustness.
- Environmental Factors: Performance was notably higher in sunlit areas compared to shaded regions, where the POD increased by nearly 6%.
Discussion
A key finding of the study was the "human baseline." A reproducibility test involving five experts revealed an agreement (F1-score) of only 0.6 to 0.7. This suggests that the automated model's performance is already approaching the limit of human consistency, as experts often disagree on exact avalanche boundaries.
Furthermore, the model's success relied heavily on spatial context. Using larger image patches (512x512 pixels) allowed the network to recognize broader topographic features like ridges and valley floors. The model also learned to prioritize the specific textures of avalanche deposits over simple brightness, which is essential for distinguishing snow features in high-glare environments.
Limitations and Future Work
Despite its success, the model faces challenges in deep mountain shadows, particularly during the winter solstice when the signal-to-noise ratio is low. Additionally, the current training data is limited to mid-winter conditions; further research is needed to determine how the model handles "wet snow" avalanches typical of the spring season.
Future improvements may include integrating "impossible" masks (e.g., excluding flat valley floors or dense forests) to reduce false positives and exploring multi-sensor fusion with Synthetic Aperture Radar (SAR) to enable monitoring through clouds and shadows.
Summary updated