Large-Scale Avalanche Mapping from SAR Images with Deep Learning-based Change Detection
Authors: Mattia Gatti, Alberto Mariani, Ignazio Gallo, Fabiano Monti
Organizations: Department of Theoretical and Applied Sciences, University of Insubria, Via Ottorino Rossi 9, Varese, 21100, Italy · Department of Science and High Technology, University of Insubria, Via Valleggio 11, Como, 22100, Italy · Alpsolut S.r.l., Via Saroch 1098/A, Livigno, 23041, Italy
Accurate change detection from satellite imagery is essential for monitoring rapid mass-movement hazards such as snow avalanches, which increasingly threaten human life, infrastructure, and ecosystems due to their rising frequency and intensity. This study presents a systematic investigation of large-scale avalanche mapping through bi-temporal change detection using Sentinel-1 synthetic aperture radar (SAR) imagery. Extensive experiments across multiple alpine ecoregions with manually validated avalanche inventories show that treating the task as a unimodal change detection problem, relying solely on pre- and post-event SAR images, achieves the most consistent performance. The proposed end-to-end pipeline achieves an F1-score of 0.8061 in a conservative (F1-optimized) configuration and attains an F2-score of 0.8414 with 80.36% avalanche-polygon hit rate under a less conservative, recall-oriented (F2-optimized) tuning. These results highlight the trade-off between precision and completeness and demonstrate how threshold adjustment can improve the detection of smaller or marginal avalanches. The release of the annotated multi-region dataset establishes a reproducible benchmark for SAR-based avalanche mapping.
Figures & tables
Figure 1: Overview of the proposed avalanche mapping workflow. SAR images (VH and VV) are preprocessed (not shown). In the training stage (left), patches and annotations are used to fit a change detection model, followed by validation and threshold tuning. During inference (right), patches are extracted from pre- and post-event SAR images and passed through the model to obtain avalanche probability maps, which are then blended into a continuous map and thresholded to produce the final binary avalanche map.
Event
Date
Extent (km 2 )
Polygons (n)
Livigno
2024-04-03
509
231
2025-01-29
251
207
2025-03-18
488
79
Nuuk
2016-04-13
226
117
2021-04-11
255
129
Pish
2023-02-21
253
113
Table 1: Avalanche inventories analyzed in this study. Date is the post‑event SAR acquisition. Extent is the planimetric area of the clipped analysis polygon (not the raster extent). Polygons is the number of mapped avalanches.
Figure 2: Geographic location of the study areas used in this work: Nuuk (a), Tromsø (b), Livigno (c), and Pish (d).
Figure 3: Example of a 128×128 input patch used for avalanche mapping. From left to right, the panels show the four SAR backscatter channels, the resampled local incidence angle (LIA) and terrain slope, and the ground-truth segmentation mask. SAR channels are independently normalized using a 2 – 98% percentile stretch. LIA and slope are visualized using the viridis and terrain colormaps, respectively.
Event
Patch size
32
64
128
Pos.
Neg.
Pos.
Neg.
Pos.
Neg.
Livigno_20240403
1353
13445
739
2919
399
480
Livigno_20250129
934
12460
561
2697
299
494
Livigno_20250318
439
25785
276
6174
162
1399
Nuuk_20160413
1366
30922
695
7265
366
1592
Table 2: Number of extracted image patches for each avalanche event, reported separately for positive (Pos.) and negative (Neg.) samples across different patch sizes (32, 64, and 128 pixels).
Figure 4: 3D schematic of the proposed architecture. Pre- and post-event SAR images are processed by weight-sharing Swin Transformer encoders to extract temporally aligned features. Auxiliary topographic inputs (slope and local incidence angle) are processed by a separate Swin encoder. The element-wise difference between the deepest pre- and post-event SAR feature maps is fused with topographic features via a residual convolutional module and decoded by a hierarchical Swin Transformer, which progressively upsamples and refines the representation to produce a binary avalanche segmentation mask. This design enables joint reasoning over spatial, temporal, and topographic cues.
Patch size
Recall
Precision
F1-score
IoU
32×32
0.7553
0.8014
0.7777
0.6362
64×64
0.7408
0.8380
0.7864
0.6480
128×128
0.7854
0.8092
0.7971
0.6608
Table 3: Performance metrics on the test set for different patch sizes with auxiliary input. Bold values indicate the best F1 and IoU scores, which were the metrics optimized during threshold tuning.
Patch size
Recall
Precision
F1-score
IoU
32×32
0.7566
0.8157
0.7850
0.6461
64×64
0.7635
0.8148
0.7880
0.6504
128×128
0.7925
0.8131
0.8027
0.6704
Table 4: Performance metrics on the test set for different patch sizes without auxiliary input. Bold values indicate the best F1 and IoU scores, which were the metrics optimized during threshold tuning.
Figure 5: Visual results for selected test patches. From left to right: VV RGB composite (pre-post-pre), VH RGB composite, predicted probability map, thresholded binary mask, and ground truth mask.
Model
Params (M)
F1
IoU
SiamUnet-diff ( Daudt et al., 2018 )
1.35
0.7370
0.5836
SiamUnet-conc ( Daudt et al., 2018 )
1.35
0.7498
0.5997
STANet ( Chen and Shi, 2020 )
2.42
0.6460
0.4771
BIT ( Chen et al., 2021 )
2.97
0.6999
0.5383
SNUNet-CD ( Fang et al., 2021 )
12.03
0.7546
0.6059
ChangeFormer ( Bandara and Patel, 2022 )
55.26
0.7720
0.6197
Table 5: Comparison of change detection models on the test set. Only F1 and IoU are reported, as F1 was used as the primary optimization target during validation. Bold indicates the best results. Parameter counts are in millions (M). Models are listed by publication date.
Size
Description
Typical Release Volume (m 3 )
1
Small
<102
2
Medium
102 – 103
3
Large
103 – 104
4
Very Large
104 – 105
5
Extremely large
>105
Table 6: Avalanche size classification according to the European Avalanche Warning Services (EAWS) standards ( European Avalanche Warning Services, 2025a ) .
Mode
Rec.
Prec.
F1
IoU
Hit %
None
0.7892
0.8059
0.7975
0.6632
62.50% (70/112)
Min
0.7710
0.8338
0.8012
0.6683
60.71% (68/112)
Max
0.8136
0.7924
0.8029
0.6707
68.75% (77/112)
Mean
0.7887
0.8211
0.8046
0.6731
64.29% (72/112)
Gaussian
0.7928
0.8199
0.8061
0.6752
65.18% (73/112)
Center Crop
0.7948
0.8171
0.8058
0.6747
65.18% (73/112)
Table 7: Performance metrics and total polygon hit rate. Best values per metric are shown in bold.
Mode
Rec.
Prec.
F1
F2
IoU
Hit %
Max (F1-opt)
0.8136
0.7924
0.8029
0.8076
0.6707
68.75% (77)
Gauss (F1-opt)
0.7928
0.8199
0.8061
0.7991
0.6752
65.18% (73)
Max (F2-opt)
0.8771
0.7021
0.7799
0.8414
0.6392
80.36% (90)
Gauss (F2-opt)
0.8633
0.7345
0.7937
0.8324
0.6579
78.57% (88)
Table 8: Metrics and total polygon hit rate with threshold tuning (F1/F2). F2 emphasizes recall four times more than precision.
Figure 6: Confusion map of predicted avalanches across the Tromsø region using the Gaussian blending strategy, generated with the threshold optimized for the F1-score.
Figure 7: Confusion map of predicted avalanches across the Tromsø region using the Max blending strategy, generated with the threshold optimized for the F2-score.
Mode
Hit-Size 2 (%)
Hit-Size 3 (%)
Hit-Size 4 (%)
Max (F1-opt)
32.00 (8/25)
74.65 (53/71)
100.00 (16/16)
Gauss (F1-opt)
28.00 (7/25)
70.42 (50/71)
100.00 (16/16)
Max (F2-opt)
64.00 (16/25)
81.69 (58/71)
100.00 (16/16)
Gauss (F2-opt)
60.00 (15/25)
80.28 (57/71)
100.00 (16/16)
Table 9: Avalanche polygon hit rates by EAWS size class (2–4).