Automated avalanche mapping from SPOT 6/7 satellite imagery: results, evaluation, potential and limitations

Hafner, Elisabeth D.; Barton, Patrick; Daudt, Rodrigo Caye; Wegner, Jan Dirk; Schindler, Konrad; Bühler, Yves

doi:https://doi.org/10.5194/tc-2022-80

Preprints

https://doi.org/10.5194/tc-2022-80

Preprints

20 Apr 2022

Status: this preprint is currently under review for the journal TC.

Automated avalanche mapping from SPOT 6/7 satellite imagery: results, evaluation, potential and limitations

Elisabeth D. Hafner^1,2,3, Patrick Barton³, Rodrigo Caye Daudt³, Jan Dirk Wegner^3,4, Konrad Schindler³, and Yves Bühler^1,2 Elisabeth D. Hafner et al.

Show author details

¹WSL Institute for Snow and Avalanche Research SLF, Davos Dorf, 7260, Switzerland
²Climate Change, Extremes, and Natural Hazards in Alpine Regions Research Center CERC¸Davos Dorf, 7260, Switzerland
³EcoVision Lab, Photogrammetry and Remote Sensing, ETH Zurich, Zurich, 8092, Switzerland
⁴Institute for Computational Science, University of Zurich, Zurich, 8057, Switzerland

Received: 05 Apr 2022 – Discussion started: 20 Apr 2022

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. This information is today still collected in a non-systematic way by observers in the field. Current research has explored and proposed applying remote sensing technology to fill this information gap by providing spatially continuous information on avalanche occurrences over large regions. Previous investigations have confirmed the high potential of avalanche mapping from remote sensed imagery to complement existing databases. Currently, the bottleneck for fast data provision from optical data is the time- consuming manual mapping. In our study we deploy a slightly adapted DeepLabV3+, a state-of-the-art deep learning model, to automatically identify and map avalanches in SPOT6/7 imagery from 24 January 2018 and 16 January 2019. We relied on 24'778 manually annotated avalanche polygons split into geographically disjoint regions for training, validating and testing. Additionally, we investigate generalization ability by testing our best model configuration on SPOT 6/7 data from 6 January 2018 and comparing to avalanches we manually annotated for that purpose. To assess the quality of the model results, we investigate the probability of detection (POD), the positive predictive value (PPV) and the F1-score. Additionally, we assessed the reproducibility of manually annotated avalanches in a small subset of our data. We achieved an average POD of 0.610, PPV of 0.668 and an F1-score of 0.625 in our test areas and found an F1-score in the same range for avalanche outlines annotated by different experts. Our model and approach are an important step towards a fast and comprehensive documentation of avalanche periods from optical satellite imagery in the future, complementing existing avalanche databases. This will have a large impact on safety related applications, making mountain regions safer.

Elisabeth D. Hafner et al.

Status: open (until 15 Jun 2022)

Post a comment Subscribe to comment alert

RC1: 'Comment on tc-2022-80', Ron Simenhois, 11 May 2022 reply

This manuscript describes an important and interesting work toward automatic avalanche detection from visual spectrum satellite imagery. Also, I am very impressed with the work to build the dataset to train, evaluate/calibrate and test. Overall, the manuscript clearly describes the methods and the results of this project. However, I have a few comments on the manuscript that need to be addressed. These are initial comments to start the discussion. I will add more comments soon.

I agree with the authors on the importance of avalanche detection from optical satellite data. Still, the authors' reasoning for using optical images is narrow and specific to the existence of a specific dataset. Are there other, more general reasoning for using optical images in addition to SAR images? I would like to see more reasoning for using optical satellite images for the avalanche field in general.

The majority of the readers of this journal are not AI specialists. I will increase this paper's readability if the authors add a short non-technical description of the DeepLad3+, what it is, and what it does.

Specific comments:

Line 78: This is the first time the term "expert" appears in the paper. Please elaborate on who the expert is and what their role is.

Lines 84, 86: Aren't POD and PPV probabilities (a number between 0 and 1)? Why are you using %?

Line 101: add a reference to ResNet

Line 101: Just out of curiosity, Chen et al. (2018) show better results with a version of Xception. Why did you decide to use the ResNet encoder?

Line 110: Doesn't it need to be Figure 2?

Reply

Citation: https://doi.org/10.5194/tc-2022-80-RC1
RC2: 'Comment on tc-2022-80', Ron Simenhois, 13 May 2022 reply

Below are more specific comments for the manuscript.
Line 125: You mention the patches' size in several places in later parts of the manuscript. Maybe mention it here (pixel and ground cover) instead (except in 4.2) for better clarity.
Lines 127-133: Your method description of balancing the data is unclear. If I understand correctly, you under-sampled non-avalanche areas by considering only parches centered around an avalanche for 95% of the images. Is this correct? Please make the description clear.
Line 129: You use "First" where is second?
Line 146: Did you try the IoU loss function to reduce the effect of the unbalanced dataset?
Lines 150-151: did transforming shadow pixels' negative values v →(−3 ・ v2) help? I don't see any reference to it in other places in the manuscript.
4.1 Results and generalization ability: This section should be reorganized for better clarity. You initially present your detection results on the test set and then introduce the previously unseen test set concept. I suggest that you swap the first and second paragraphs.
Table 2: some of the fields are in bold font. Why? Am I missing something?
Figure 6: For clarity, I recommend replacing "Avalanche Score" with the commonly used term: "model confidence"
Line 232: Do you think that the experts' difficulty with low light areas may cause lower quality data in shadowy areas and, as a result, reduce the model performance in these areas?
Table 4: Is this pixel-level comparison or detection of an avalanche?
Figure 8: The color scale here is unclear. What do the 1 to 5 values mean?
4.4 Limitations of this study section: Is this it? How about the experts' calcification variation in dark areas that suggests that the data quality may be unclear?
Line 285: The F1-scores seem low… Is this increase in an F1-score? It doesn't add up otherwise... Please make it clear.

Reply

Citation: https://doi.org/10.5194/tc-2022-80-RC2

Elisabeth D. Hafner et al.

Viewed

Total article views: 301 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
223	68	10	301	3	1

HTML: 223
PDF: 68
XML: 10
Total: 301
BibTeX: 3
EndNote: 1

Views and downloads (calculated since 20 Apr 2022)

Month	HTML	PDF	XML	Total
Apr 2022	124	32	5	161
May 2022	99	36	5	140

Cumulative views and downloads (calculated since 20 Apr 2022)

Month	HTML	PDF	XML	Total
Apr 2022	124	32	5	161
May 2022	99	36	5	140

Viewed (geographical distribution)

Total article views: 286 (including HTML, PDF, and XML) Thereof 286 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 24 May 2022

Short summary

Knowing where avalanches occur is very important information for several disciplines, for example avalanche warning, hazard zonation or risk management. Satellite imagery can provide such data systematically over large regions. In our work we propose a machine learning model to automize the time- consuming manual mapping. Additionally, we investigate expert agreement for manual avalanche mapping, showing that our network is equally good as the experts in identifying avalanches.


Total:	0
HTML:	0
PDF:	0
XML:	0