Snow avalanches are natural hazards, occurring in snow
covered mountain terrain worldwide. Present avalanche
research and forecasting relies on complete avalanche activity records in
a given area over an entire winter season, which cannot be
provided with traditional, mainly field based
methods. Remote sensing, using
weather, and light independent SAR satellites has the
potential of filling these data gaps,
however, to date their use was limited by high
acquisition costs, long repeat
cycles, and small ground swath.
Sentinel-1A (S1A), on the other
hand, operational since October 2014 provides
free-of-charge,
20 m spatial resolution, 250 km
Snow avalanches (hereafter called avalanches), cause about 250 fatalities annually, as well as high economic costs due to mitigation measurements and loss of infrastructure (Schweizer, 2008). Avalanche research, is primarily hazard research, focusing on the understanding of avalanche formation in space and time and the resulting consequences. Traditionally, field-testing of snow properties, field reconnaissance of avalanche activity and modelling of both, research the avalanche problem. This approach is handicapped by high-risk exposure and observational bias towards easily accessible areas. Creating a complete avalanche activity record, which is important for forecasting, mitigation and research is thus not possible over an entire winter in any given area.
Remote sensing can potentially fill this data gap by providing safe, unbiased observational data on avalanche activity over fixed areas. To date, ground, air-, and space borne SAR sensors seem to be the most useful tools due to their high to very high resolution, as well as weather and light independency (Bühler et al., 2014). However, their operational use is limited by data availability, high acquisition costs, small ground swaths and long repeat passes.
However, the Sentinel-1A SAR satellite, launched in April 2014, provides free-of-charge, high-resolution SAR data with large ground swath and short repeat time since October 2014. In this short communication, we show for the first time that is possible to detect avalanche debris in Sentinel-1A images. This is an important result for future operational use of SAR data in detecting avalanches over large areas and therefore for creating a complete avalanche record.
The detectable part of an avalanche is its debris, being the snow mass that an avalanche entrains in its slide path. In this paper, we use both avalanche and avalanche debris to describe the same, detectable feature.
A quantitative electromagnetic model of backscatter from avalanche debris is not published. However, Eckerstorfer and Malnes (2014) provided a quantitative interpretation of relative backscatter from dry and wet avalanche debris. Based on the theory of backscatter from undisturbed dry and wet snowpack (Ulaby et al., 1986), increased snow volume, liquid water content, snow density and surface roughness in avalanche debris leads to increased backscatter. They further assume that the backscatter from the air–snow surface interface due to a rough snow avalanche debris surface is the dominant snow parameter that allows for detection of avalanche debris (Eckerstorfer and Malnes, 2014).
The first space-borne SAR detection of avalanches was done by Wiesmann et al. (2001) using ERS1/2 data. A change detection algorithm was applied, utilizing the change in backscatter between two SAR images. Avalanche debris appeared as tongue-shaped features with increased backscatter and sharp backscatter contrast to the surrounding. More recently, Malnes et al. (2013) Eckerstorfer et al. (2014) and Eckerstorfer and Malnes (2014) used both change detection algorithms, as well as single image detection to detect avalanches in Radarsat-2 Ultrafine Mode (RS-2 U) images. Eckerstorfer and Malnes (2014) showed that it is possible to quantify the magnitude of an avalanche cycle in a forecasting region using satellite-borne SAR. However, spatially comprehensive avalanche detection was not achievable, due to the inconsistent availability of RS-2 U data.
From the evening of 24 December until 27 December 2014,
there was a significant snowfall (at least 50
Sentinel-1A has a repeat cycle of 12
days. The satellite provides on regular basis 4–5 partial/full coverage over the area we study
(counties of Troms and Nordland, Northern Norway)
(Fig. 1, green rectangle). The
Sentinel-1 IW mode images (Interferometric wide swath) have
a spatial resolution of 20
The Sentinel-1A (S1A) images were downloaded in the
standard format ground range detected high-resolution
(S1-GRDH) which are focused SAR
images, georeferenced to a flat earth ellipsoid with 20
Manual avalanche detection in single backscatter images is possible
(Eckerstorfer and Malnes,
2014), however, enriched RGB
image composites display changes in backscatter in
colour, improving manual detection
significantly. A RGB composite image (Wiesmann et al., 2001) includes the reference
image in the R and B channels, and the current image
(with avalanche activity) in the G channel:
[R, G, B]
In order to validate the SAR interpretations we have collected in-situ data from avalanche debris in the valley Lavangsdalen, easily accessible from the E8 road. Using a handheld GPS, we marked parts of the furthest runout of two wet slab avalanches, that released from the east-facing valley side.
Additionally we took photos of the avalanches, both in the valley Lavangsdalen, and Breivikeidet, confirming that the detected features were really avalanches.
In a descending path S1A image from 28 December 2014 (not
shown), no significant avalanche activity is
visible. This drastically changes in the ascending S1A
picture from 6 January 2014, which shows a snapshot of
avalanche activity from the avalanche cycle of 1–3 January 2015 (Fig. 1). The entire S1A ground swath
(
These avalanches cluster in areas with steep topography along the
fjords. There is a distinct peak of avalanche activity
from the town Tromsø towards the southeast (Fig. 1). However, there is also a cluster
of avalanche activity in the interior of the county Troms (Fig. 2b). Given the synoptic meteorological condition and the
instabilities in the snowpack we expected more avalanche activity in the
Lofoten–Vesterålen area than we were able to
observe in the S1A image, however the rugged
topography, creating significant radar shadow and
foreshortening effects, strongly limit the manual
detection of avalanche debris (Fig. 2c). Despite the
limitation of these radar artefacts, manual detection of
avalanche debris is for an experienced observer,
relatively straightforward. A strong backscatter
contrast, assumed to be of the order of
2–3
With a spatial resolution of 20
In Fig. 3 we compare a Radarsat-2 Ultrafine image
from 3 January 2015, with a S1A image from 6 January 2015 (similar image that is used in the RGB composite in Figs. 1 and 2). In both images, avalanche debris
is detectable as light grey features due to increased
backscatter. We have used this manual detection method
in single RS2-U images successfully to create a record
of avalanche activity in the county of Troms in March 2014 (Eckerstorfer and
Malnes, 2014). In the
RS2-U image, 102 avalanches were
detectable within the
We drove through the valley Breivikeidet (Fig. 2a), close to Tromsø, 23 January 2015 to take pictures of the detected avalanche debris. A frequently used road leads to Breivikeidet, making field observations easy and efficient.
The valley Lavangsdalen is an avalanche-prone location, where the main road E8 is recurrently endangered from both valley sides (Fig. 3). We collected GPS tracks from two avalanche debris on 8 January 2015. In Fig. 5 we present a multi-sensor, multi-temporal series of SAR images, all from ascending paths, from Lavangsdalen. In the S1A image from 27 December 2014, no avalanche activity is visible. The two features with high backscatter are natural debris flow tracks. Figures 5b, 4c and d, consistently depict avalanche activity that occurred after the avalanche cycle. The collected GPS tracks agree very well with the edge of the avalanche features in all images. This gives us high confidence that we are detecting avalanche debris in the S1A images.
We further took pictures of the avalanche debris for further field validation of the detected avalanche debris, presented in Fig. 6.
Avalanche debris detection in satellite-borne SAR images.
Satellite-borne SAR data has the advantage of weather and light independent image acquisition. This is especially advantageous at northern latitudes, such as the location of our study area, where the Polar night (the sun is below the horizon for two months) largely limits the use of optical data during the winter months.
The clear contrast in backscatter, between avalanche debris with high backscatter and surrounding, undisturbed snowpack with lower backscatter enabled the relatively straightforward visual detection of avalanche debris. The lack of an electromagnetic backscatter model for avalanche debris (disturbed snowpack), as well as the lack of appropriate parameterization of snow parameters in avalanche debris, limits the quantification of the backscatter difference. However, given the comparison between RS2-U and S1A images, as well as the field verification of some detected avalanches, we are confident that we are detecting avalanches. A sharp, clearly delimited increase in backscatter from avalanche debris is due to increased snow depth, snow water equivalent and most importantly increased surface roughness. Thus, we assume that the largest backscatter contribution, which is enhanced under wet snow conditions, stems from surface scattering at the air–snow interface. The total backscatter of wet snow avalanche debris is, however, smaller than that of dry snow avalanche debris (Eckerstorfer and Malnes, 2014; Ulaby et al., 1986).
By creating RGB composite images, using a change detection algorithm of backscatter between two images of the same geometry but acquired at different times, manual avalanche debris detection is further enhanced. The visibility of the backscatter contrast between avalanche debris and surrounding snowpack is enhanced by the stark colour contrast, which enables detection and interpretation also from less experienced observers. This method was until now limited by the availability of reference images with similar geometry, when for example using RS2-U images. There are, however, limitations that become especially apparent in very steep, rugged alpine topography (Fig. 2c). The presence of radar shadows, layover and foreshortening in many areas make avalanche debris detection much more difficult. There are also natural features that have a similar geometry to avalanche debris (e.g. small glaciers, debris from rock fall, debris flows), as well as a strong backscatter contrast to their surroundings. Misinterpreting features as avalanche debris, as well as under detection are problems we cannot quantify yet. We believe however, that more avalanche debris is missed due to radar artefacts than features that have accidentally been interpreted as avalanche debris.
Comparing avalanche debris detection in very high resolution RS2-U and high-resolution S1A images, we find that RS2-U images are technically better suited to the task of detection due to their higher spatial resolution. This allows for detection of small avalanches that are not visible in S1A images. All the large, significant avalanches with run-outs close to a road are, however, clearly detectable in the S1A image. Moreover, acquiring both ascending and descending path S1A images allows for avalanche debris detection at all aspects, at least in the area that is covered by both geometries (i.e. alpine terrain according to the Sentinel-1 high-level operation plan of ESA 2014).
In a recent ESA report, Bühler et al. (2014) identified potential users of remote sensing
data in avalanche forecasting. Amongst the users were
national and regional avalanche warning services that have high
willingness-to-pay for auxiliary
data provided by remote sensing, due to
a large, inaccessible forecasting area and a sparse
observational network. One of the main identified data
gaps was “avalanche activity”. Requested were hourly
temporal resolution, approximately 10
While S1A products cannot provide near-real time
avalanche activity data, they can provide data with
a 12 day repeat interval. The main
advantage over terrestrial radar systems is the large area
covered, which is
Such a complete avalanche activity record is not only of interest for avalanche warning services, but also critical in avalanche research. Such complete datasets are of great importance for statistical avalanche forecasting (Buser, 1983; Eckerstorfer and Christiansen, 2011; Föhn et al., 1977; Hendrikx et al., 2014), to return-period calculations (Eckert et al., 2010), identification of avalanche hazard on infrastructure (Hendrikx and Owens, 2008; Margreth et al., 2003), and increasingly climate change related studies (Fitzharris and Schaerer, 1980; Marty and Meister, 2012).
For these applications however, all avalanche debris needs to be detected and recorded in a geospatial database. However, the accumulation of large datasets may become problematic for the analysis and storage of data. Manual detection of avalanche debris in a handful of S1A scenes every 12 days is labour intensive and time consuming. Thus, an automatic avalanche debris detection algorithm is needed. Such a detection algorithm could utilize the backscatter contrast, as well as the typical morphological features of avalanche debris. By further identifying and neglecting terrain where avalanche debris is unlikely to occur (based on slope angle, dense forest cover), the algorithm would have less pixels to search for, which would reduce computational problems.
In this study, we show for the first
time, that Sentinel-1A images in
interferometric wide swath mode (IW) (
We are confident that avalanche detection using S1A images as a first critical step towards operational use of SAR data in avalanche forecasting. This is mainly due to the free-of-cost availability of at least two images per 12 day repeat cycle, the weather and light independency, and the large ground swath. Thus, significant avalanche activity can be monitored throughout a winter in a large forecasting region. However, manual detection is time consuming, thus automatic avalanche detecting algorithms need to be improved, in order to automate the entire processing chain.
This research is financed by RDA Troms (competence centers in and for northern environments). Radarsat-2 images were acquired under the Norwegian RS-2 agreement. Sentinel-1A data was acquired from the Sentinel-1 Scientific Data Hub.
Overview of acquired SAR images, both S1A and RS2 images.
S1A RGB image composite, where two S1A images with similar ascending path are merged. The reference image is from 13 December 2014, visualized in the green channel, the image from 6 January 2015 is visualized in the red and blue channel. Avalanche debris appears as green, tongues-shaped features; the furthest runout location is marked with a green point. The purple colour indicates a change in physical snow parameters, most likely a change from dry to wet snow. The three red rectangles depict areas of interest, shown in Fig. 2.
Three areas of interest from the S1A
RGB image composite in Fig. 1
(red rectangles). Avalanche debris appears as
green, tongues-shaped
features; the furthest runout location is marked
with a green point. The purple colour indicates a change
in physical snow parameters, most likely a change from
dry to wet snow. The numbered avalanches in Breivikeidet
valley in
Comparison
between a Radarsat-2 Ultrafine
(RS2-U) image from 3 January 2015
Fieldwork validation of detected avalanche debris in the valley Breivikeidet, 23 January 2015. Avalanche debris was still clearly visible then, as the period since New Year was dry and cold. The numbers correspond to the detected avalanche debris in Fig. 2a.
Multi-sensor and multi-temporal series of SAR images from Lavangsdalen valley, containing ascending path S1A and RS2-U images. The road E8 through Lavangsdalen valley is in orange. The collected GPS tracks of two avalanche debris are visualized with red lines.
Fieldwork validation of detected avalanche debris in the valley Lavangsdalen, 8 January 2015. Avalanche debris was still clearly visible then, as the period since New Year was dry and cold. The numbers correspond to the detected avalanche debris in Fig. 5b.