💻 computer science

Large-Scale Avalanche Mapping from SAR Images with Deep Learning-based Change Detection

This paper presents a deep learning-based, end-to-end pipeline for large-scale avalanche mapping using bi-temporal Sentinel-1 SAR imagery, which achieves high detection performance across multiple alpine regions and establishes a reproducible benchmark dataset for future research.

Original authors: Mattia Gatti, Alberto Mariani, Ignazio Gallo, Fabiano Monti

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mattia Gatti, Alberto Mariani, Ignazio Gallo, Fabiano Monti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to find where a massive snow avalanche has just happened. The problem? You can't go there in person. The weather is too stormy, the terrain is too dangerous, and it's often pitch black or covered in clouds.

This paper is about building a super-smart digital detective that can look at satellite photos and say, "Aha! An avalanche happened right here!"

Here is the story of how they built this detective, explained in simple terms with some fun analogies.

1. The Camera: Seeing Through the Storm

Usually, satellites take pictures using light (like a regular camera). But if it's cloudy or night, the camera is blind.

The Solution: The researchers used Sentinel-1, a special satellite that uses radar (like a bat using echolocation). It shoots invisible waves at the ground.
The Analogy: Imagine trying to find a wet spot on a carpet in a dark room. You can't see it with your eyes, but if you tap the floor, the wet spot sounds different than the dry spot. The radar "taps" the ground. When snow slides down a mountain, it changes the texture of the ground (making it rougher or smoother). The radar "hears" this change, even through clouds and darkness.

2. The Detective's Training: Two Photos vs. Three

The researchers wanted to know: Does our detective need extra help to solve the case?

The "Extra Help" (Multimodal): Traditionally, experts give the computer a map of the mountain's slope and direction (like giving a detective a map of where crimes usually happen).
The "Pure Detective" (Unimodal): The researchers tried training the AI using only the "before" and "after" radar photos. No maps, no extra data. Just the raw evidence.
The Surprise: They found that the Pure Detective was actually better! It learned to spot the changes just by looking at the two photos. Adding the extra maps didn't help much and actually made the system more complicated. It's like realizing you don't need a GPS to find your way home if you just have a really good memory of the streets.

3. The Puzzle Pieces: Tiling the Map

The satellite photos are huge—like a giant mural of the entire Alps. The computer's brain (the GPU) isn't big enough to look at the whole mural at once.

The Solution: They cut the giant photo into small puzzle pieces (patches), like 128x128 pixel squares.
The Overlap: They made the pieces overlap slightly, like shingles on a roof. This ensures that if an avalanche is right on the edge of a piece, the computer still sees it clearly.
The Blending: When the computer solves each puzzle piece, it has to glue them back together. The researchers tested different ways to glue them:
- Average: Taking the middle opinion of all overlapping pieces.
- Max: If any piece thinks it's an avalanche, they say "Yes!" (Very sensitive).
- Gaussian: A fancy way of giving more weight to the center of the piece and less to the edges.
- The Winner: The Gaussian method (giving weight to the center) worked best, creating a smooth, seamless map without weird seams.

4. The "Goldilocks" Threshold: How Strict Should We Be?

Once the computer makes a map, it has to decide: "Is this really an avalanche?" It uses a "confidence score."

The F1-Score (The Conservative Detective): This setting tries to be perfect. It only flags an avalanche if it's 100% sure.
- Result: Very few false alarms, but it might miss small or tricky avalanches.
The F2-Score (The Cautious Detective): This setting cares more about not missing anything. It's okay to have a few false alarms if it means catching every single avalanche.
- Result: It catches 80% of the avalanches (including tiny ones), but it also flags a few areas that looked like avalanches but weren't.
The Lesson: If you are a safety officer trying to save lives, you want the F2-Score (catch everything, even if you have to double-check the false alarms later). If you just want a clean map for a report, the F1-Score is better.

5. The Big Result

Speed: This system is incredibly fast. It can scan the entire Alps (a huge area) in less than 40 minutes on a powerful computer.
Accuracy: It found avalanches with an accuracy score of about 80%, which is excellent for such a difficult task.
The Gift: The researchers didn't just build the tool; they released the dataset (the training photos and answers) to the public. This is like giving every other detective in the world the same practice exam so they can all get better.

Summary

This paper proves that you don't need a complicated, multi-tool system to find avalanches from space. A simple, smart AI that just looks at "before and after" radar pictures can do the job better than complex systems. By tweaking how strict the AI is, we can choose between a "clean map" or a "safety-first map" that catches even the smallest slides.

It's a giant leap forward for keeping people safe in the mountains, using the power of AI to see what human eyes cannot.

1. Problem Statement

Snow avalanches pose significant threats to life, infrastructure, and ecosystems in mountainous regions. Traditional monitoring relies on field observations, which are limited by poor visibility, harsh weather, and terrain inaccessibility. While Sentinel-1 Synthetic Aperture Radar (SAR) offers all-weather, day/night monitoring capabilities, automated avalanche mapping remains challenging due to:

Ambiguity: Backscatter changes caused by avalanches can be confused with snowfall, wind redistribution, melt-refreeze cycles, and other surface changes.
Scalability: Existing workflows often rely on complex, manually engineered rules or auxiliary terrain data (e.g., Digital Elevation Models) that increase preprocessing complexity and limit large-area deployment.
Data Scarcity: There is a lack of large-scale, manually validated datasets for benchmarking deep learning models in this specific domain.

The core research question is whether a unimodal approach (using only pre- and post-event SAR imagery) can outperform or match multimodal approaches (which include terrain-derived inputs like slope and local incidence angle) in terms of accuracy and scalability.

2. Methodology

A. Data and Preprocessing

Data Source: Sentinel-1 C-band SAR imagery (VV and VH polarizations) from four distinct alpine regions: Livigno (Italy), Nuuk (Greenland), Pish (Tajikistan), and Tromsø (Norway).
Ground Truth: A manually validated inventory of avalanche polygons derived from field observations, reports, and site photos.
Preprocessing: Performed using SNAP software, including radiometric calibration, terrain correction, speckle filtering, and conversion to decibel (dB) scale.
Input Handling:
- Auxiliary Data: Local Incidence Angle (LIA) and terrain slope were computed from a 30m DEM and upsampled to match SAR resolution.
- Normalization: SAR values were clipped to a valid range (−40 to 20 dB). Invalid pixels (e.g., radar shadow) were replaced with a sentinel value derived from the normalized distribution to preserve patch structure without needing a separate mask channel.
- Patch Generation: Overlapping patches (32×32, 64×64, 128×128 pixels) were extracted to increase sample diversity and enable seamless inference blending.

B. Model Architecture

Base Architecture: A Multi-branch Swin Transformer V2 (Tiny variant).
Design:
- Dual Encoders: Two weight-shared encoders process pre-event and post-event SAR images to learn temporal change representations.
- Auxiliary Encoder: A separate encoder processes LIA and slope (used in multimodal experiments).
- Fusion: Feature maps from the deepest encoder stage are differenced (post minus pre) and concatenated with topographic features. A residual convolutional module fuses these representations before a shared decoder produces the binary segmentation mask.
Training Strategy:
- Loss Function: Weighted Binary Cross-Entropy ( $w_p=3.0$ ) to address severe class imbalance (few avalanche pixels vs. vast background).
- Sampling: Balanced sampling ensures equal numbers of positive and negative patches per epoch.
- Augmentation: Geometric (flips, rotations) and radiometric (Gaussian noise, gain perturbation) augmentations.

C. Inference and Post-Processing

Blending Strategies: To handle large scenes, overlapping tile predictions are fused using several methods: Per-pixel averaging, Max/Min blending, Gaussian-weighted blending, and Center cropping.
Threshold Tuning:
- F1-Optimized: Maximizes the F1-score (balance of precision and recall).
- F2-Optimized: Maximizes the F2-score (weights recall 4x higher than precision) to prioritize event detection over false positive control.
Morphological Refinement: A closing operation (dilation followed by erosion) is applied to remove isolated noise and fill small gaps.

3. Key Contributions

Systematic Comparison of Modalities: The study demonstrates that a unimodal SAR-only formulation achieves more consistent and often superior performance compared to multimodal models incorporating terrain data. This simplifies the pipeline and reduces computational overhead.
Open Benchmark Dataset: The authors released a manually validated, multi-region avalanche inventory (geodatabase) covering four distinct geographic areas, establishing a reproducible benchmark for SAR-based avalanche mapping.
Operational Threshold Analysis: The paper provides a detailed analysis of the precision-recall trade-off, showing how adjusting thresholds (F1 vs. F2) and blending strategies can tailor the system for different operational needs (e.g., conservative mapping vs. high-recall screening).

4. Key Results

Performance Metrics

Best Configuration: Using 128×128 patches with Gaussian blending and F1-optimized thresholding.
- F1-Score: 0.8061
- IoU: 0.6752
- Polygon Hit Rate: 65.18% (defined as detecting ≥50% of the avalanche area).
Unimodal vs. Multimodal: Removing the auxiliary terrain branch (LIA/Slope) resulted in virtually identical or slightly improved performance (F1 increased from 0.7971 to 0.8027 on the test set), proving that the deep network can learn necessary spatial/radiometric cues directly from SAR data.
Baseline Comparison: The proposed Swin-UNet architecture outperformed established change detection baselines (e.g., SiamUnet, STANet, ChangeFormer, TinyCD), achieving the highest F1 (0.8027) and IoU (0.6604) on the test set.

Thresholding and Recall

F2-Optimized Tuning: Switching to an F2-optimized threshold significantly improved recall for smaller avalanches.
- Polygon Hit Rate: Increased from ~65% (F1-opt) to 80.36% (F2-opt).
- Size 2 Avalanches: Hit rate nearly doubled (from ~28% to ~60%).
- Trade-off: Precision decreased (from ~0.82 to ~0.73), indicating more false positives in snow accumulation zones, which is acceptable for screening workflows.

Efficiency

Scalability: The system processes approximately 87 km²/s on an NVIDIA A100 GPU. The entire Alps (~200,000 km²) could be processed in under 40 minutes.

5. Significance and Conclusion

This work establishes that deep learning-based change detection using only bi-temporal SAR imagery is sufficient for large-scale avalanche mapping, challenging the conventional wisdom that terrain data is strictly necessary.

Practical Impact: The proposed end-to-end pipeline offers a scalable, reproducible solution for operational avalanche monitoring. By releasing the dataset and code, the authors enable rigorous benchmarking and future research.
Operational Flexibility: The study highlights that there is no single "optimal" setting; rather, the system can be tuned based on the use case:
- F1-Optimized: Best for high-precision, conservative mapping where false positives must be minimized.
- F2-Optimized: Best for emergency screening where missing an event (false negative) is more critical than reviewing extra candidates.
Future Directions: The authors suggest integrating physical constraints (e.g., avalanche simulations) during post-processing to filter false positives in physically implausible areas and investigating scene-level models to capture broader contextual information.