Elisabeth Hafner

Snow avalanches are a major natural hazard in mountainous areas, posing a significant risk to lives and infrastructure. Knowing the location, frequency, and magnitude of past snow avalanche occurrences is vital to mitigate these risks. Avalanche mapping is therefore critical in risk mitigation. Previous research has explored various techniques to automate snow avalanche mapping from remote sensing sources, including satellite synthetic aperture radar (SAR), optical airborne, and satellite imagery. However, employing historical optical data sources for full snow avalanche delineation has received less attention. These datasets, created in the analog era, could shed light on the patterns of past avalanche activity. The winter of 1999 in the European Alps is notable for its significant avalanche activity with extreme avalanche runouts. It is well documented through aerial images covering over 12,000 square kilometers in the Swiss Alps and neighboring Austria. Our study leverages deep learning techniques to map the release, path, and deposition areas of visible snow avalanches in these historical images. Building upon previous work in deep learning for automated mapping with optical SPOT 6/7 satellite imagery, we have applied an avalanche segmentation model to historical data. The model is based on a convolutional neural network and relies on digital elevation data and orthorectified optical imagery to produce a pixel-level binary snow avalanche segmentation mask. This task requires the adaption of the methods, originally developed and trained on multi-spectral satellite data, for application on grey-scale data, a problem referred to as a domain gap. We do this by augmenting the historical training set with contemporary samples. The contemporary data is transformed to match the target data domain using a Generative Adversarial Network and coupled with the historical data to train the model. We also offer insights into the estimated uncertainty associated with labels and predictions, serving as a resource for future dataset users. Our research shows how cutting-edge mapping techniques can be adapted to historical imagery.

Humans have been exposed to snow avalanches ever since they inhabited or travelled through mountainous regions. For centuries, ways of managing risk and mitigating damage due to avalanches have been limited. For example, forests prevented the release of avalanches, protecting the villages below. In the past ca. 150 years, avalanche protection measures and risk reduction strategies have become increasingly sophisticated. With a growing population, the rise of alpine tourism, and major transportation corridors running through the mountains, dealing with avalanches has become more important than ever. Operational services such as the avalanche warning service, hazard mapping, or the installation of protection and mitigation measures for endangered zones are important tools for avalanche risk management. The efficiency of these tools depends on knowledge about past avalanche occurrences. In this thesis, novel methods are developed to automatically provide this information over large regions. Specifically, state-of-the-art deep learning technology is used to extract such information from optical satellite and webcam imagery. The avalanches required to train the deep learning models are identified and mapped by human experts in the domain. Therefore, the thesis also focuses on the consistency of the avalanche area identified by various domain experts.

In the first part of the thesis, a DeeplabV3+ model is adapted to automatically identify and map avalanches from optical SPOT 6/7 satellite imagery (1.5 m resolution). The model is trained, validated, and tested with more than 24'000 manually annotated avalanche polygons. The data originate from two avalanche periods in January 2018 and January 2019 and cover an area of more than 22'000 km2. In addition, the quality of the model and the reproducibility of avalanches manually annotated by experts are assessed for a small subset of the data.

The second part of the thesis investigates in more detail the reproducibility of estimates of avalanche dimensions by human experts in three user studies. The first study analyzes the classification of ten avalanches into five standardized size categories by each of 170 avalanche experts from Europe and North America. The second and the third study examine avalanches manually mapped from oblique photographs (6 avalanches, 10 participants) or from remotely sensed imagery (2.9 km², 5 participants), respectively.

The third part of the thesis leverages interactive avalanche segmentation (IAS) to combine human expert knowledge with deep learning. Here, when mapping avalanches from webcam imagery, the user collaborates with the previously trained model. The use of the model is supposed to make avalanche mapping more accurate and efficient. For this purpose, we adapt a state-of-the-art interactive segmentation model based on HRNet+OCR and train it for avalanche segmentation from webcam imagery. The human user interacts with the model through confirming or corrective feedback.

In summary, this thesis makes a substantial contribution to the development of an operational automatic avalanche mapping service. The thesis provides the first automatic avalanche mapping from optical satellite imagery with deep learning, it quantifies, for the first time to this extent, the reproducibility of human avalanche estimates, and it presents a first interactive approach for the mapping of avalanches from webcam imagery. Thereby, this thesis contributes to a more efficient use of data and better provision of information on past avalanche occurrences, and thus it benefits decision making in safety-relevant applications.

For many safety-related applications such as hazard mapping or road management, well-documented avalanche events are crucial. Nowadays, despite the variety of research directions, the available data are mostly restricted to isolated locations where they are collected by observers in the field. Webcams are becoming more frequent in the Alps and beyond, capturing numerous avalanche-prone slopes. To complement the knowledge about avalanche occurrences, we propose making use of this webcam imagery for avalanche mapping. For humans, avalanches are relatively easy to identify, but the manual mapping of their outlines is time intensive. Therefore, we propose supporting the mapping of avalanches in images with a learned segmentation model. In interactive avalanche segmentation (IAS), a user collaborates with a deep-learning model to segment the avalanche outlines, taking advantage of human expert knowledge while keeping the effort low thanks to the model’s ability to delineate avalanches. The human corrections to the segmentation in the form of positive clicks on the avalanche or negative clicks on the background result in avalanche outlines of good quality with little effort. Relying on IAS, we extract avalanches from the images in a flexible and efficient manner, resulting in a 90 % time saving compared to conventional manual mapping. The images can be georeferenced with a mono-photogrammetry tool, allowing for exact geolocation of the avalanche outlines and subsequent use in geographical information systems (GISs). If a webcam is mounted in a stable position, the georeferencing can be reused for all subsequent images. In this way, all avalanches mapped in images from a webcam can be imported into a designated database, making them available for the relevant safety-related applications. For imagery, we rely on current data and data archived from webcams that cover Dischma Valley near Davos, Switzerland, and that have captured an image every 30 min during the daytime since the winter of 2019. Our model and the associated mapping pipeline represent an important step forward towards continuous and precise avalanche documentation, complementing existing databases and thereby providing a better base for safety-critical decisions and planning in avalanche-prone mountain regions.

Systematic mapping of avalanches using remote-sensing technologies has become a reliable and – when combined with machine-learning algorithms – also time-effective way to document avalanche activity over large areas. Making use of these technological advances, we compared the avalanche activity for three periods with widespread activity of very large and extremely large avalanches, when the highest danger level 5 (very high) was forecast for large parts of the Swiss Alps. In these three periods (in 1999, 2018, 2019), avalanches were captured with different remote-sensing instruments: in 1999, with airplane imagery, in 2018 and 2019 with multispectral satellite data. Comparing the activity of all three periods to Large Scale Hazard Indication Modelling (LSHIM) of avalanches with a 100-year return period allowed to compare how much of the potential avalanche terrain was active and to quantitatively compare activity between these periods. We found that 19% of the avalanche release areas showed activity in 1999 compared to 16% in 2018 and 8% in 2019. Considering the avalanche area delimited using the 100-year return period, 17% were active in 1999, 11% in 2018 and 6% in 2019. By comparing the three avalanche periods, we contribute to better understand the extent and return periods of very large and extremely large avalanches (size 4 and 5) and demonstrate, how avalanche mapping from remotely sensed data can make avalanche databases more complete.

Am 14. April 2024 um 06:56 Uhr ereignete sich in der Westflanke des Piz Scerscen (Berninamassiv, Engadin, Graubünden) auf einer Höhe von 3640 m ü. M. ein grosser Bergsturz. Rund 5.5 Millionen m³ Fels und Gletschereis lösten sich, stürzten auf den darunter liegenden Vadret da Tschierva und rissen in der Sturzbahn eine dicke Schneedecke und grosse Mengen Gletschereis mit sich. Dieses Gemisch aus Gestein, Eis und Schnee floss über das Gletscherplateau, durch die darunterliegenden Eisfälle und zwischen den Seitenmoränen des Gletschers hinunter und erzeugte eine grosse Staubwolke, bevor es sich auf rund 2100 m ü. M. im Val Roseg, fast 6 km vom Anriss ablagerte. Das Ereignis forderte glücklicherweise keine Opfer und verursachte keine Infrastrukturschäden. In diesem Bericht beschreiben wir die Bedingungen, die zum Zeitpunkt des Ereignisses herrschten und nutzen die verfügbaren Informationen, um die Prozesse zu analysieren, die vor und während des Ereignisses stattfanden.

Accurate snow cover maps are an important tool in a large variety of applications including avalanche research. The WSL Institute for Snow and Avalanche Research (SLF) has developed a snow cover mapping system for the Dischma valley in Davos, based on ground-based cameras. Their goal has been to validate snow-cover maps derived from satellite data. In the currently implemented system, several steps require manual work. The goal of this paper is to automate SLF's approach by applying a deep learning framework. For the training of our models, we have access to data from multiple cameras mounted in the Dischma valley. In our approach, we first apply deep learning to perform fog classification and then to do pixel-wise snow segmentation. Unlike the current procedure our method is independent of image- and camera-specific thresholds. In our experiments, we compare a model that is trained on images from all cameras to camera-specific models for both tasks. For the fog classification, we aim for a higher recall to be able to detect most of the foggy images. We show that, when tuned to achieve the same precision as the baseline, our model improves recall by 17%. For the snow segmentation, we evaluate our models based on the Fl-score. When using our fully automated machine learning model without any manually selected thresholds, we achieve a higher Fl-score only for a part of the cameras in the valley, and on average, this comes at a cost of 2.3%. We also show that for the case of a new camera being mounted in the Dischma valley, our models for both tasks can still be used for that new camera and will achieve similar results compared to the currently mounted cameras.

Information on snow depth and its spatial distribution is important for numerous applications, including natural hazard management, snow water equivalent estimation for hydropower, the study of the distribution and evolution of flora and fauna, and the validation of snow hydrological models. Due to its heterogeneity and complexity, specific remote sensing tools are required to accurately map the snow depth distribution in Alpine terrain. To cover large areas (>100 km²), airborne laser scanning (ALS) or aerial photogrammetry with large-format cameras is needed. While both systems require piloted aircraft for data acquisition, ALS is typically more expensive than photogrammetry but yields better results in forested terrain. While photogrammetry is slightly cheaper, it is limited due to its dependency on favourable acquisition conditions (weather, light conditions). In this study, we present photogrammetrically processed high-spatial-resolution (0.5 m) annual snow depth maps, recorded during the peak of winter over a 5-year period under different acquisition conditions over a study area around Davos, Switzerland. Compared to previously carried out studies, using the Vexcel UltraCam Eagle Mark 3 (M3) sensor improves the average ground sampling distance to 0.1 m at similar flight altitudes above ground. This allows for very detailed snow depth maps in open areas, calculated by subtracting a snow-off digital terrain model (DTM, acquired with ALS) from the snow-on digital surface models (DSMs) processed from the airborne imagery. Despite challenging acquisition conditions during the recording of the UltraCam images (clouds, shaded areas and fresh snow), 99 % of unforested areas were successfully photogrammetrically reconstructed. We applied masks (high vegetation, settlements, water, glaciers) to increase the reliability of the snow depth calculations. An extensive accuracy assessment was carried out using check points, the comparison to DSMs derived from unpiloted aerial systems and the comparison of snow-free DSM pixels to the ALS DTM. The results show a root mean square error of approximately 0.25 m for the UltraCam X and 0.15 m for the successor, the UltraCam Eagle M3. We developed a consistent and reliable photogrammetric workflow for accurate snow depth distribution mapping over large regions, capable of analysing snow distribution in complex terrain. This enables more detailed investigations on seasonal snow dynamics and can be used for numerous applications related to snow depth distribution, as well as serving as a ground reference for new modelling approaches and satellite-based snow depth mapping.

Monitoring snow depth is important for applications such as hydrology, energy planning, ecology, and safety evaluation for outdoor winter activities. Most methods able to estimate snow depth for large regions can only do so in a spatial resolution of up to 1 km ground sampling distance (GSD). This limits their usage in high alpine areas, where this resolution fails to capture local snow distribution patterns caused by the pronounced topographical features. In this work we use a recurrent convolutional neural network to estimate snow depth at high spatial resolution (10 m GSD), weekly, and at large scale based on satellite data sources and elevation maps, without the need for measurement stations on the ground. The proposed method achieves unprecedented results for large-scale, high-resolution snow depth mapping. The resulting maps are evaluated over a period of three years against high-fidelity snow depth maps obtained with airborne photogrammetry. Finally, we also produce well-calibrated uncertainty estimates for every individual snow depth estimate via a probabilistic regression framework.

Consistent estimates of avalanche size are crucial for communicating not only among avalanche practitioners but also between avalanche forecasters and the public, for instance in public avalanche forecasts. Moreover, applications such as risk management and numerical avalanche simulations rely on accurately mapped outlines of past avalanche events. Since there is not a widely applicable and objective way to measure avalanche size or to determine the outlines of an avalanche, we need to rely on human estimations. Therefore, knowing about the reliability of avalanche size estimates and avalanche outlines is essential as errors will impact applications relying on this kind of data. In the first of three user studies, we investigate the reliability in avalanche size estimates by comparing estimates for 10 avalanches made by 170 avalanche professionals working in Europe or North America. In the other two studies, both completed as pilot studies, we explore reliability in the mappings of six avalanches from oblique photographs from 10 participants and the mappings of avalanches visible on 2.9 km² of remotely sensed imagery in four different spatial resolutions from 5 participants. We observed an average agreement of 66 % in the most frequently given avalanche size, while agreement with the avalanche size considered “correct” was 74 %. Moreover, European avalanche practitioners rated avalanches significantly larger for 8 out of 10 avalanches, compared to North Americans. Assuming that participants are equally competent in the estimation of avalanche size, we calculated a score describing the factor required to obtain the observed agreement rate between any two size estimates. This factor was 0.72 in our dataset. It can be regarded as the certainty related to a size estimate by an individual and thus provides an indication of the reliability of a label. For the outlines mapped from oblique photographs, we noted a mean overlapping proportion of 52 % for any two avalanche mappings and 60 % compared to a reference mapping. The outlines mapped from remotely sensed imagery had a mean overlapping proportion of 46 % (image resolution of 2 m) to 68 % (25 cm) between any two mappings and 64 % (2 m) to 80 % (25 cm) when compared to the reference. The presented findings demonstrate that the reliability of size estimates and of mapped avalanche outlines is limited. As these data are often used as reference data or even ground truth to validate further applications, the identified limitations and uncertainties may influence results and should be considered.

Ereigniskataster aufgetretener Lawinen sind eine wichtige Grundlage für die Gefahrenkartierung. Diese dient der Umsetzung verschiedener Instrumente im integralen Risikomanagement, wie zum Beispiel der Gefahrenzonenplanung. Die dokumentierten Ereignisse werden ausserdem vom Lawinenwarndienst im Nachhinein für eine Validierung ihrer Prognose verwendet. Die derzeit durch beobachtende Personen aufgenommenen Lawinen sind räumlich unvollständig und konzentrieren sich um häufig begangenes Gebiet sowie auf Schadlawinen. Ausserdem ist die Qualität des Ereigniskatasters kaum verifizierbar bezüglich der Qualität der Umrisse und der Vollständigkeit der Ereignisse, da schlechte Sicht während einer Lawinenperiode die zeitliche Erfassung verzögern kann und Lawinen durch Neuschnee und Wind überdeckt werden können. Satellitenbilder bieten die Möglichkeit, Lawinen über grosse Gebiete, unabhängig von der Zugänglichkeit, zeitnah zu dokumentieren. Für zwei Lawinenperioden haben wir dies mit optischen SPOT6/7 Satellitenbildern für eine Gesamtfläche von ~22 000 km² gemacht und daraus manuell 24 778 Lawinen kartiert. Mit einer Untersuchung basierend auf Fotos vom Boden und aus dem Helikopter als Referenzdaten, konnten wir zeigen, dass diese Erfassung in Bezug auf die Vollständigkeit zuverlässig ist. Um die Lawinenerkennung schneller und wiederholbar zu machen, haben wir sie mit künstlicher Intelligenz, genauer gesagt Deep Learning, automatisiert. Die Möglichkeit grossflächig und automatisiert Lawinen zu erfassen ist ein wichtiger Schritt zu vollständigeren und zuverlässigeren Lawinendatenbanken.

Cadastres of avalanches that have occurred are an important basis for hazard mapping. This is used to implement various instruments in integral risk management, such as hazard zone planning. In addition, the documented events are used by the avalanche warning service to validate their forecast in retrospect. The avalanches currently recorded by observers are spatially incomplete and concentrated around accessible areas and damaging avalanches. In addition, the quality of the avalanche cadastre is difficult to verify in terms of the quality of the outlines and the completeness of the events, since poor visibility during an avalanche period can delay the recording and avalanches can be covered through fresh snow and wind.
Satellite imagery allows for the quick documentation of avalanches over large areas, regardless of accessibility. For two avalanche periods, we used SPOT6/7 optical satellite imagery for a total area of ~22,000 km² to manually map 24,778 avalanches. With a study based on ground and helicopter photos as reference data, we showed that this mapping is reliable in terms of completeness. To make avalanche mapping faster and repeatable, we automated it with artificial intelligence, more precisely Deep Learning. The ability to detect avalanches on a large scale and in an automated way is an important step towards more complete and reliable avalanche databases.

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 the application of 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 remotely 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 SPOT 6/7 imagery from 24 January 2018 and 16 January 2019. We relied on 24 778 manually annotated avalanche polygons split into geographically disjointed 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 it 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.

In January 2018, heavy snowfalls caused an extraordinary avalanche situation in Switzerland with an extent not observed since the avalanche winter of 1999. The amount of precipitation was far above the long-term average und in parts of the Swiss Alps even twice the average. For the first time since 1999, the highest danger level 5-Very High was forecast for a 36-hour period and for most regions of the Swiss Alps. We present key findings of the analysis of this extraordinary event and show that the mitigation efforts taken after the catastrophic events in February 1999 were effective. Despite very high avalanche activity, there was no loss of life and damage to property was limited.

Information on the occurrence and size of snow avalanches is important for the avalanche warning, decision making for the closure of traffic lines as well as for hazard mapping and the planning of mitigation measures. Today such information is limited to a small fraction of wellobserved regions and time periods when visual observations are possible.
Recent advances in the mapping of avalanche events based on satellite data have great potential to close this gap as large areas can be covered in a spatial continuous manner. This includes regions, which are inaccessible for observers. In this paper we present an overview on current investigations on the potential and performance of satellite mappings (optical and SAR) to complete existing avalanche databases and expand them to previously unobserved regions.

The mitigation of rapid mass movements involves a subtle interplay between field surveys, numerical modelling, and experience. Hazard engineers rely on a combination of best practices and, if available, historical facts as a vital prerequisite in establishing reproducible and accurate hazard zoning. Full-scale field tests have been performed to reinforce the physical understanding of debris flows and snow avalanches. Rockfall dynamics are - especially the quantification of energy dissipation during the complex rock-ground interaction - largely unknown. The awareness of rock shape dependence is growing, but presently, there exists little experimental basis on how rockfall hazard scales with rock mass, size, and shape. Here, we present a unique data set of induced single-block rockfall events comprising data from equant and wheel-shaped blocks with masses up to 2670 kg, quantifying the influence of rock shape and mass on lateral spreading and longitudinal runout and hence challenging common practices in rockfall hazard assessment.

The spatial distribution and size of avalanches are essential parameters for avalanche warning, avalanche documentation, mitigation measure design and hazard zonation. Despite its importance, this information is incomplete today and only available for limited areas and limited time periods. Manual avalanche mapping from satellite imagery has recently been applied to reduce this gap achieving promising results. However, their reliability and completeness have not yet been verified satisfactorily.
In our study we attempt a full validation of the completeness of visually detected and mapped avalanches from optical SPOT 6, Sentinel-2 and radar Sentinel-1 imagery. We examine manually mapped avalanches from two avalanche periods in 2018 and 2019 for an area of approximately 180 km² around Davos, Switzerland, relying on ground- and helicopter-based photographs as ground truth. For the quality assessment, we investigate the probability of detection (POD) and the positive predictive value (PPV). Additionally, we relate our results to conditions which potentially influence avalanche detection in the satellite imagery.
We statistically confirm the high potential of SPOT for comprehensive avalanche mapping for selected periods (POD = 0.74, PPV = 0.88) as well as the reliability of Sentinel-1 (POD = 0.27, PPV = 0.87) for which the POD is reduced because mainly larger avalanches are mapped. Furthermore, we found that Sentinel-2 is unsuitable for the mapping of most avalanches due to its spatial resolution (POD = 0.06, PPV = 0.81). Because we could apply the same reference avalanche events for all three satellite mappings, our validation results are robust and comparable. We demonstrate that satellite-based avalanche mapping has the potential to fill the existing avalanche documentation gap over large areas, making alpine regions safer.

Accurate and timely information on avalanche occurrence is key for avalanche warning, crisis management and avalanche documentation. Today such information is mainly available at isolated locations provided by observers in the field. The achieved reliability, considering accuracy, completeness and reliability of the reported avalanche events, is limited. In this study we present the spatially continuous mapping of a large avalanche period in January 2018 covering the majority of the Swiss Alps (12 500 km²).
We tested different satellite sensors available for rapid mapping during the first avalanche period. Based on these experiences, we tasked SPOT6 and SPOT7 for data acquisition to cover the second, much larger avalanche period. We manually mapped the outlines of 18 737 individual avalanche events, applying image enhancement techniques to analyze regions in the shade as well as in brightly illuminated ones. The resulting dataset of mapped avalanche outlines, having unique completeness and reliability, is evaluated to produce maps of avalanche occurrence and avalanche size. We validated the mapping of the avalanche outlines using photographs acquired from helicopters just after the avalanche period.
This study demonstrates the applicability of optical, very high spatial resolution satellite data to map an exceptional avalanche period with very high completeness, accuracy and reliability over a large region. The generated avalanche data are of great value in validating avalanche bulletins, in completing existing avalanche databases and for research applications by enabling meaningful statistics on important avalanche parameters.