A method for automated snow avalanche debris detection through use of synthetic aperture radar (SAR) imaging

Transcription

1 PUBLICATIONS Earth and Space Science RESEARCH ARTICLE Key Points: Radar remote sensing of snow avalanche debris Automatic detection algorithm Operational monitoring of avalanches A method for automated snow avalanche debris detection through use of synthetic aperture radar (SAR) imaging H. Vickers 1, M. Eckerstorfer 1, E. Malnes 1, Y. Larsen 1, and H. Hindberg 1 1 Norut Tromsø, Tromsø, Norway Correspondence to: H. Vickers, Hannah.Vickers@norut.no Citation: Vickers, H., M. Eckerstorfer, E. Malnes, Y. Larsen, and H. Hindberg (2016), A method for automated snow avalanche debris detection through use of synthetic aperture radar (SAR) imaging, Earth and Space Science, 3, , doi:. Received 8 MAR 2016 Accepted 22 SEP 2016 Accepted article online 29 SEP 2016 Published online 23 NOV 2016 Abstract Avalanches are a natural hazard that occur in mountainous regions of Troms County in northern Norway during winter and can cause loss of human life and damage to infrastructure. Knowledge of when and where they occur especially in remote, high mountain areas is often lacking due to difficult access. However, complete, spatiotemporal avalanche activity data sets are important for accurate avalanche forecasting, as well as for deeper understanding of the link between avalanche occurrences and the triggering snowpack and meteorological factors. It is therefore desirable to develop a technique that enables active mapping and monitoring of avalanches over an entire winter. Avalanche debris can be observed remotely over large spatial areas, under all weather and light conditions by synthetic aperture radar (SAR) satellites. The recently launched Sentinel-1A satellite acquires SAR images covering the entire Troms County with frequent updates. By focusing on a case study from New Year 2015 we use Sentinel-1A images to develop an automated avalanche debris detection algorithm that utilizes change detection and unsupervised object classification methods. We compare our results with manually identified avalanche debris and field-based images to quantify the algorithm accuracy. Our results indicate that a correct detection rate of over 60% can be achieved, which is sensitive to several algorithm parameters that may need revising. With further development and refinement of the algorithm, we believe that this method could play an effective role in future operational monitoring of avalanches within Troms and has potential application in avalanche forecasting areas worldwide The Authors. This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made. 1. Introduction Snow avalanches (from hereon called avalanches), both natural and human released, occur throughout the winter season in the mountainous regions of Troms County in northern Norway (Figure 1). In addition, there has been a significant increase in recreational use of the mountains during wintertime by both skiers and snowmobile drivers, which has in turn resulted in an increase in avalanche fatalities especially in the last few years. Between 2004 and 2009, an average of 3 fatalities occurred per winter season, whereas this number had increased to an average of 8.5 over the five winter seasons between 2009 and 2014 ( Moreover, the occurrence of avalanches can lead to road closures and community evacuations and thus high economic costs. To mitigate the avalanche risk, public avalanche warning has been operational in northern Norway since January 2013 [Engeset, 2013]. Avalanche risk assessment is dependent on complete, spatiotemporal data sets of avalanche activity in a given region [McClung and Schaerer, 2006]. Avalanche activity is a good indicator of avalanche danger, especially during periods of considerable to high avalanche danger level [Schweizer, 2003]. With traditional field-based monitoring, a complete database of avalanche occurrences is not achievable. This not only limits evaluation and validation of the avalanche danger but also restricts a deeper understanding of the link between avalanche release and the triggering snowpack and meteorological conditions. Field-based monitoring of avalanche activity uses guidelines to estimate morphological and topographical parameters of avalanches from a safe distance [Greene et al., 2010]. However, field-based monitoring is both time and labor intensive and limited by weather, light conditions, and avalanche danger. Radar remote sensing of avalanche debris on the other hand is a relatively new technique [Brogioni et al., 2014] which has been shown to be effective for avalanche detection over large spatial areas and does not rely on specific weather or light conditions [Eckerstorfer et al., 2016]. The potential of synthetic aperture radar (SAR) images for identifying avalanches was first demonstrated by Wiesmann et al. [2001] where it was shown that the increased backscatter in spaceborne ERS 1/2 C band SAR data resulting from rough, compacted avalanche debris snow could be utilized to detect a single avalanche. ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 446

2 Figure 1. Geographical overview of two overlapping Sentinel-1A scenes from an ascending pass on 6 January (green) and descending pass on 9 January 2015 (yellow) that cover partly all nine forecasting regions in northern Norway. Case studies are presented from Lavangsdalen and Tamokdalen (red boxes) and the green and yellow points illustrate manually identified avalanche debris for the two dates, totaling 707 avalanche debris. Recent studies by Malnes et al. [2013], Bühler et al. [2014], and Eckerstorfer and Malnes [2015] have made use of change detection methods in order to detect avalanche debris. This technique identifies the temporal change in radar backscatter between two SAR scenes acquired, one from before avalanche activity has taken place and another from after. As avalanche debris exhibits an increase in radar backscatter, they become manually identifiable as elongated, tongue-shaped, and downslope-extending features. Using RADARSAT-2 ultrafine images at 3 m resolution, Eckerstorfer and Malnes [2015] manually identified avalanche debris, reporting a median backscatter difference within avalanche debris of approximately 3.7 db. The authors also showed that the median backscatter difference from nonavalanche homogeneous snow in a 500 m buffer region surrounding these debris was 1.5 db, with the distributions of the two populations overlapping at 1.5 db. Based on their observations, the authors also formulated a qualitative model for radar backscatter from avalanche debris based on earlier work as described by Ulaby et al. [1986]. It is suggested that the combination of increased snow volume, water content, density, and surface roughness could lead to stronger scattering of the radar wave with respect to scattering from homogeneous undisturbed snow surrounding the debris, thus making the debris detectable by SAR. However, manual avalanche debris identification, whether performed using optical or radar-based observations, is time consuming and limited by observator bias. In the context of building up a substantial database of avalanche activity, there is therefore a need for an automated detection algorithm. Attempts to develop automated methods for avalanche detection have been made by Bühler et al. [2009] using optical remote sensing data collected by the ADS40 instrument for a test site at Davos in Switzerland. In their methodology, they construct an image of the Normalized Difference Angle Index, a measure of surface roughness, to which they apply texture analysis and object-based classification. The authors reported an accuracy of 87% for correct detections (with respect to ground truth) but highlight that the method is prone to misclassifications of small features such as wind-modeled or artificial snow and ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 447

3 snow-covered vegetation, which all exhibit similar texture to avalanche debris. They also highlighted that avalanche debris that becomes covered after a new snowfall is no longer detectable by this method. A similar approach based on object-based image interpretation was more recently implemented by Lato et al. [2012], also using optical remote sensing imagery and ecognition software which makes use of spatial and spectral information in order to classify images acquired for avalanche-prone areas in Western Norway and Davos, Switzerland. They validated their results against manually generated avalanche maps and reported an accuracy of over 0.9. Larsen et al. [2011] have also reported on detection of avalanches in high resolution satellite optical imagery whereby two texture segmentation methods have been demonstrated and evaluated. They have investigated use of grey level coherence matrices and directional filters in order to segment images by snow/vegetation type. However, the technique described is not fully automatic since it makes use of training data in order to generate a model database from which the segmentation algorithm can identify different class types. Use of training samples, which does not include all the possible avalanche types that can exist in an image, was found to be a weakness of this method, leading to misclassification of pixels. High temporal resolution, ground-based SAR images have also been used to develop automated avalanche detection algorithms [Martinez-Vasquez and Fortuny Guasch, 2008] with the objective of aiding near real-time monitoring and warning of avalanche activity. In their work, images are obtained at 12 min intervals, thereby facilitating the use of differential coherence as a parameter in detecting avalanche release. Their detection processor consists of three fundamental stages, namely, segmentation (thresholding), features extraction, and classification. Based on a supervised subset of more than 60,000 images, the authors report that the detection processor had an accuracy of 60%, which was shown to be sensitive to the degree of filtering and thresholding applied. While there exists a variety of approaches that have been explored in order to develop automated or semiautomated procedures for avalanche detection, automatic detection of avalanche debris in spaceborne SAR images has to our knowledge, not been fully achieved yet. In this work, we have attempted to develop an automated avalanche detection algorithm by making use of Sentinel-1A images at 20 m resolution together with a K-means classification procedure. K-means is an unsupervised clustering method, which divides data into clusters that are as compact as possible [Theodoridis and Koutroumbas, 2006]. K-means clustering has also been used widely in other applications including oil spill detection [Migliaccio et al., 2015; Wang et al., 2014; Ganesan, 2015] and sea ice classification [Gill et al., 2013; Karvonen, 2010; Korosov et al., 2015] and is therefore a well-developed approach for classification applications using remote sensing data. We have deliberately chosen to implement an unsupervised classification procedure since it is our ultimate goal to develop a fully automated, operational system for avalanche debris detection that does not require additional human input. Alternative supervised methods, while potentially more accurate, rely on selection of training samples in order to identify data classes. The use of such methods in an avalanche debris detection application would therefore be dependent on (manual) selection of training sites, which is in turn subject to human interpretation of the remote sensing data. We envisage that the algorithm could be a useful tool for long-term monitoring of avalanche activity in all 23 avalanche forecasting regions in Norway that are monitored by the Norwegian Avalanche Centre [www. varsom.no]. As Sentinel-1 data are available worldwide, our technique could be applied to any avalanche forecasting region worldwide that covers large areas and lacks field observations of avalanche activity. 2. Data Sources and Processing 2.1. Sentinel-1A SAR Images Sentinel-1A was launched in April 2014 and has been operational since October The interferometric wide swath mode (IW) covers a swath of approximately 250 km and is extended often up to 1000 km. The swath length varies according to a fixed observation pattern that is set by European Space Agency (ESA)/ Copernicus as part of the operational monitoring programme. The repeat time for identical image geometry is currently 12 days, but due to the satellites polar orbit, high-latitude areas can be observed more frequently (every 4 to 5 days) at multiple viewing geometries, which is ideal for monitoring purposes. The IW mode has 20 m 5 m resolution (azimuth-ground range), and images are acquired in two polarizations (bands), copolarization ( VV ) and cross polarization ( VH ). We have processed all images from both VV and VH polarizations. ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 448

4 Table 1. Summary of Sentinel-1A Images Utilized in the Case Study Areas Path Number Avalanche Debris Image Reference Image 131 (ASC) 2015/01/ /12/ (ASC) 2015/01/ /12/ (DES) 2015/01/ /12/16 The geocoding process is done using the GSAR software [Larsen et al., 2005]. Here we use the calibrated Sentinel-1 IW GRDH product (Ground Range Detected High resolution), which is sampled with 10 m pixel spacing on the ellipsoid. The following geocoding steps are (1) calibration using an annotated lookup table (from the GRDH product); (2) multilooking, 2 2 pixels averaging to suppress speckle; (3) map the selected Universal Transverse Mercator (UTM)-projected output grid to the radar coordinates using a 10 m digital elevation model produced by the Norwegian Mapping Authority ( and precision orbits vectors available from the European Space Agency (ESA) ( and (4) project the backscatter product to the output grid using the radar coordinate mapping (from step 3) and cubic interpolation. The output radar backscatter image (σ 0 ) is stored as a GeoTIFF using the UTM zone 33 N, WGS-84 projection and sampled to 20 m pixel spacing. This ensures a sufficiently high multilook factor (~20 looks) to suppress most SAR speckle artifacts. The geocoding process also generates mask files for radar shadow and layover caused by the radar geometry and the terrain Data Selection For the purpose of developing an algorithm for the automatic detection of avalanche debris, we have used a selection of Sentinel-1A images that were acquired in the week following an avalanche cycle, which took place around New Year 2015 in Troms County. A summary of the images used is given in Table 1. An intense low-pressure system affected large parts of northern Norway, leading to a rise in air temperatures of up to almost 6 C, strong westerly winds and heavy rainfall. This resulted in substantial, natural wet snow avalanche activity, especially in the entire county of Troms (Figure 1). A large number of avalanche debris were manually identifiable in the first set of Sentinel-1A backscatter images acquired on 6, 8, and 9 January 2015 after the avalanche cycle took place. A total of 505 and 404 avalanche debris was manually interpreted in the entire swath of the ascending (1 January 2015) and descending (9 January 2015) images, respectively. Figure 1 illustrates the overlap of two of the image acquisitions from 6 and 9 January 2015, here shown as RGB (Red, Green, Blue) composite images. Green and yellow dots indicate the locations of avalanche debris identified manually on 6 (ascending pass) and 9 January (descending pass), respectively, and the red boxes highlight the two case study areas of interest in this paper, Lavangsdalen and Tamokdalen. Using an image with a lot of observable avalanche debris is beneficial in the latter stages of the algorithm development when we wish to compare the results of the automatic detection with those that have been manually identified. We also note that while the snow cover was wet during the avalanche cycle, it was refrozen and dry again during the actual days of image acquisition. Dry and wet snow scatter the radar signal differently [Ulaby et al., 1986]. In the case of dry snow, the radar wave penetrates the snow volume, with the majority of the backscattered signal originating from the snow/ground interface and smaller scattering contributions from the snow volume and air/snow interface. For wet snow the radar wave does not penetrate to the snow/ground interface and it is only the air/snow surface from which the backscattered radar wave originates. Additionally, snow temperature affects the backscattering coefficient of the snow. For example, it has been experimentally shown for ground-based radars operating at C, X, and Ku band that snow at lower temperature gives higher backscatter than snow at higher temperatures [Baumgartner et al., 1999]. Therefore, we would expect a decrease in radar backscatter in the image where the snow condition was wet during the period with higher (air) temperatures, relative to an image where the snow condition was dry and cold. In the automatic avalanche debris detection algorithm, we make use of change detection and unsupervised classification methods to identify avalanche debris-type features automatically. In order to implement change detection we therefore require a suitable reference image before a given avalanche event in addition to the image containing avalanche debris so that a difference image can be formed. The reference image must have the same satellite viewing geometry as the avalanche debris (hereon debris ) image and should have been acquired not too long before the debris image (preferably 12 days before if available). This usually ensures that the background conditions, i.e., the nonavalanche areas of the image, appear the same in both ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 449

5 Earth and Space Science images, such that the difference between the two images in the background is close to zero and the avalanche debris therefore appear more distinct due to their higher backscatter difference. Figure 2. (a) Single backscatter Sentinel-1A image for Tamokdalen recorded on 13 December 2014 in VV polarization. This image acts as a reference image without any avalanche debris. (b) Single backscatter Sentinel-1A image for Tamokdalen recorded on 6 January 2015 in VV polarization. This image has the same geometry and ascending path (131) as in Figure 2a, acting as the debris image after avalanche activity has taken place. (c) RGB composite image formed from the avalanche debris image (6 January 2015) and the reference backscatter image (13 December 2014) for VV polarization. Avalanche debris appears as bright green features; layover and shadows are masked out in white. ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION We show an example of the Tamokdalen area (case study 1) in Figure 2 where a large number of avalanche debris were identified. Here the reference Sentinel-1A backscatter image recorded on 13 December 2014 is shown in Figure 2a and the avalanche debris image acquired on 6 January 2015 is shown in Figure 2 b. Air temperatures measured at the Tamokdalen observations site (230 m above sea level) show that the mean temperature on 13 December 2014 was 7.5 C while on 6 January 2015 the mean temperature was 10.5 C. No precipitation was recorded on either of the two dates, indicating similar meteorological conditions would have been present in the two images. An RGB (Red, Green, Blue) composite image for this acquisition, which is produced by combining the reference and debris images, is shown in Figure 2c. In order to construct the RGB image, the reference image acquired on 13 December 2014 forms the red and blue channels while the debris image containing the avalanche debris forms the green channel. An increase in backscatter appears green, while a decrease in backscatter appears purple and areas of no temporal change appear grey. The avalanche features, visible in the debris image in Figure 2b, appear as bright green features in the RGB image in Figure 2c, which indicate an increase in backscatter with respect to the reference image. It is these features that we wish to detect automatically without any prior knowledge of the area where the debris is present Manual Avalanche Identification For the purpose of comparing manual and automated methods, a manual 450

6 identification of avalanche debris in the RGB composite images from 6, 8, and 9 January 2015 was carried out by an avalanche expert with knowledge of SAR image interpretation, avalanche activity in space and time and topography of the area of interest. This was done by marking out, using ArcGIS 10, the edges of green, tongue-shaped, downslope-extending features that are typical characteristics of avalanche debris in RGB images. To avoid misinterpretation of other natural features with high backscatter (e.g., glaciers, snow fields, and lakes), a slope mask was used in order to distinguish avalanche from nonavalanche terrain [Bühler et al., 2009]. We realize that manual identification of avalanche debris is subject to observer bias, even though the change detection method makes misclassification highly unlikely. Limitations are set by the image resolution in detecting small avalanche debris, as well as in the accurate delineation of the avalanche debris, especially the delineation of the uphill part. While the terminus of the avalanche debris is in the vast majority of cases easily identifiable, delineation of where the sliding path ends and the avalanche debris starts is difficult. When comparing manually identified and automatically detected avalanche debris pixels, overestimation or underestimation can occur. 3. Automated Detection Algorithm The Sentinel-1A image covering the Tamokdalen area at a spatial resolution of 20 m is composed of approximately1.1 millionpixels. Inorderto maximizetheprobability ofdetectingpixelsthatmaypotentiallycorrespond to avalanche debris while also making the process as computationally efficient as possible, we break down the image analysis into smaller windows of pixels (1 km 1 km). In Figure 3 we illustrate the individual steps in the form of a flow chart, which we will frequently refer to in this section in order to describe the process. For each window, a backscatter difference, or change detection image, is formed by subtracting the reference image from the avalanche image (Figure 3a). We subsequently carry out a masking step (Figure 3b) whereby pixels located within areas of radar layover or shadow are masked out from further processing since we cannot detect avalanches there. This mask is automatically generated during the geocoding stage of the image processing. We also make use of a vegetation map to identify and mask out water bodies, in particular sea and fjords that do not freeze but may be affected by waves that can lead to high backscatter, which may incorrectly be classified as potential avalanche debris. Lastly, a digital elevation model is used to calculate the gradient at each pixel in the window. Since avalanche debris is mainly observed in the runout zone where the terrain gradient is generally less than 35 [Bühler et al., 2009], we identify all pixels which correspond to terrain steeper than 35 in order to construct a slope mask. A potential pitfall of applying this slope threshold is that debris resulting from smaller avalanches that can stop on steeper slope angles may be masked out by the algorithm. A final combined mask is constructed using pixels corresponding to radar layover/shadow, water, and terrain steeper than 35. All pixels in this combined mask are excluded from the search for avalanche debris pixels. Once the masking step has been carried out, we determine the percentage of pixels remaining in the masked backscatter difference image (Figure 3c) whose backscatter difference value exceeds a specified threshold value. This step acts as an initial screening of the difference data so that only windows with potential avalanche debris undergo further processing. We can thus potentially reduce the area over which the automatic detection algorithm is applied and the computational time required to process the entire area. In the current version of the algorithm, we require that greater than 1% of all remaining (unmasked) pixels have a backscatter change that is larger than 6 db. This corresponds to a minimum detectable area of ( ) = 25 pixels in the 1 1 km window even if no pixels are masked out. If the pixels in the current data window meet this criterion, we then proceed with the data processing. We further preprocess the backscatter difference images acquired in both VV and VH polarizations by applying a 3 3 median filter. Features extraction is performed on the two filtered images (Figure 3d) that are subsequently used as input features to a K-means clustering procedure (Figure 3e). In the algorithm the K-means clustering procedure is initialized from a pair of randomly selected pixels and calculates the total dissimilarity between all pixels and the class representatives. A total of 30 initializations are performed, and the classification that results in the smallest dissimilarity is selected. We use two classes since we expect a pixel in the final masked backscatter difference image to fall either into the avalanche or not avalanche category. The algorithm is written such that class 1 is associated with the highest data values, while those associated with lower values are assigned class 2. ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 451

7 Figure 3. Flow chart summarizing the data processing chain used in the automatic avalanche detection algorithm. We have deliberately chosen to apply a higher backscatter difference threshold to the data in order to be confident that only windows containing avalanche debris undergo further processing and classification. This is necessary since one of the major drawbacks with K-means classification is the sensitivity to the selected number of classes. The algorithm will separate the available data into two classes regardless of the actual separability of the data. Therefore, if we choose a lower threshold we might classify windows containing only homogeneous snow or other nonavalanche features into the avalanche and no avalanche classes. The output of the K-means procedure is an array containing only the class number, and we cannot automatically determine whether a class assignment of 1 corresponds to avalanche debris or simply pixels that had higher backscatter difference relative to the other pixels in the same window. We record all class 1 pixels and transfer these results onto a DEM corresponding to the window area (Figure 3f). The procedure is repeated for every window in the Sentinel-1A image until the entire image has been masked, filtered, and classified. A final 5 5 median filter is applied to the avalanche map to remove any remaining noise/speckle-like features. 4. Results 4.1. Case Study 1: Tamokdalen In case study 1 we show the results of object classification for the Tamokdalen area using the Sentinel-1A backscatter difference image from 6 January In this case, we have chosen the image acquired on 13 December 2014 to be the reference image since this image was acquired with identical viewing geometry and before the period during which both heavy snowfall and avalanche activity began to take place. Figure 4 shows a digital elevation model for Tamokdalen with the final avalanche pixels overlaid using the scenes acquired on 6 January In this figure the pink-colored regions indicate the automatically detected avalanche pixels and the green solid outlines represent the manually identified avalanche debris outlines that were constructed from visible inspection of the RGB image composed for the backscatter difference image from 6 January. Figure 4 shows that of the 70 individual avalanche debris features that were manually identified from the RGB image, 40 of them (57%) were also captured by the automatic detection algorithm. In terms of the number of avalanche pixels, 62% were identified both manually and automatically at the same locations, which is roughly in agreement with the number of detected features. Although the automatically detected avalanche debris does not have the exact shape of the manually identified debris, we see that the location and size of the automatically detected debris generally agree. The correctly detected avalanche debris appears to include both smaller and larger sized debris. However, we can see that in general, the avalanche debris that were missed by the automatic detection algorithm seem to correspond to smaller-sized avalanches. As noted earlier, the slope threshold of 35 may unintentionally exclude debris that result from smaller avalanches that stop on steeper slope angles. This may be a possible explanation for the somewhat poor detection of the smaller avalanches in this case, along with the limiting ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 452

8 Figure 4. Comparison between manually identified (green outlines) and automatically detected avalanche debris (pink areas) for Tamokdalen. Both manual identification and automatic detection were carried out in the difference image formed from combining the debris image from 6 January 2015 with a reference image from 13 December 2014, both with similar geometry from ascending path 131. spatial image resolution. The algorithm has also identified potential avalanche debris that was not initially marked out manually. We discuss these differences in a more quantitative fashion in section 4.3. Figure 5 shows a similar set of results as for Figure 4 but for the Sentinel-1A image acquired on 9 January The backscatter difference image for this day was constructed using a reference image from 16 December We see that the automatic detection performed slightly poorer on 9 January debris image compared with 6 January 2015 in relation to the number of debris features that were manually identified, even though the correct automatic detections represent well the size and morphology of the manual identifications. In total, 53 individual avalanche debris were identified manually. Of these, the automatic detection algorithm detected only seven. When we inspect weather data from the Tamokdalen observation site, it can be noted that there was an increase in the measured snow depth from 43 cm on 6 January to 48 cm on 9 January 2015 and most probably larger increase at higher elevations. Thus, it could be speculated that the additional snow that fell on top of the New Year avalanche debris may have led to less favorable detection conditions in terms of the data characteristics and thus poorer performance of the automatic detection algorithm. In order to complete the set of observations acquired following the New Year avalanche cycle we combine the detection maps for the three viewing geometries that were available on 6, 8, and 9 January Figure 6 illustrates on a DEM for Tamokdalen all pixels that were automatically detected as avalanche debris from these 3 days. The pixels have been color coded according to the frequency of detection in the set of three images, where pink-colored pixels correspond to detection in only one of the three images, blue-colored pixels indicatedetections in twoimages,and green pixelsarethose thathave beendetected onallthreedates. Theuse of the three geometries is beneficial since some areas may be masked out as a result of radar layover/shadow on the ascending passes (6 and 8 January) but not necessarily on the descending pass (9 January) and vice versa. In total, 6823 avalanche pixels were detected on 6 January, rising to 8415 by including 8 January detections that were not detected at the same locations as on 6 January. A total of 8938 pixels were identified to be avalanche debris when all three detection maps were combined. This corresponds to a 30% increase in the number of detected avalanche pixels compared with the detection results for only 6 January 2015 and 68% correct detections with respect to the manually identified avalanche pixels carried out on the 6 January backscatter difference image. Hence, there was only an increase of 6% (from 62%) in the number of correct detections achieved by combining the three detection maps. Therefore, even though the combination of all three ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 453

9 Earth and Space Science Figure 5. Comparison between manual (blue outlines) and automatically detected avalanche debris (red areas) for Tamokdalen. Both manual and automatic detection were carried out in the debris image from 9 January 2015 with reference image from 16 December 2014, both with similar geometry from descending path 168. detection maps resulted in an overall increase in the number of avalanche debris detected, it did not help substantially to detect pixels that had originally been missed from the first image on 6 January Case Study 2: Lavangsdalen Results of applying the automatic detection algorithm to the second area of interest, Lavangsdalen (Figure 1), are shown in Figure 7 for the classified Sentinel-1A image acquired on 6 January For this Figure 6. Combined automatic detection results using avalanche maps produced for Sentinel-1A acquisitions on 6, 8, and 9 January 2015 ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 454

10 Earth and Space Science Figure 7. Comparison between manual (red outlines) and automatically detected avalanche debris (blue areas) for Lavangsdalen. Both manual and automatic detection were carried out in the debris image from 6 January 2015 with reference image from 13 December 2014, both with similar geometry from ascending path 131. The green rectangle depicts avalanche debris we validated in the field (Figure 8). case study we have used the same reference image acquired on 13 December 2014 as an input to the automatic detection algorithm. Qualitatively and quantitatively, the results are similar to those presented for Tamokdalen in Figure 4. The automatically detected avalanche debris are shown by the blue regions, which are overlaid on a digital elevation model for Lavangsdalen. Using a threshold of 6.0 db, a reasonable majority (28) of all the manually detected avalanche debris (39) or 72% are found by the detection algorithm, but the exact shape and size of the detected debris do not completely match that of the debris which were marked out manually, as indicated by the red outlines. Also similar to the results for Tamokdalen, the detection algorithm finds additional backscatter features that have been classified as avalanche debris. The debris that were not captured from the backscatter difference image using the automatic detection algorithm can be seen to be associated with smaller areas of avalanche debris. As for the Tamokdalen case study, the differences between the manual and automatic detections will be dealt with in further detail in section 4.3. In Figure 8 we show an example of avalanche debris that was manually identified and automatically detected. We have furthermore visited the site in the field and collected a GPS track by walking along the perimeter of the avalanche debris, as well as a field photograph from the opposing hillslope (Figure 8c) where an outline of the avalanche debris has been marked out manually to aid interpretation by an observer. A qualitative comparison of the actual avalanche debris outline with the GPS track and the manual interpretation of the backscatter show very high agreement. Both the flanks and the terminus of the avalanche debris are identified correctly. The manual interpretation resulted in 365 pixels covered by avalanche debris, compared to 379 pixels detected as avalanche debris by the automatic avalanche debris detection algorithm. Both populations intersect with 321 pixels. ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 455

11 Figure 8. (a) Avalanche debris in a RGB composite image formed from the avalanche debris image (6 January 2015) and the reference backscatter image (13 December 2014) for VV polarization. GPS track, manual interpretation, and automatic detection of the avalanche debris outline are visualized. (b) RADARSAT-2 Ultrafine mode single backscatter image (3 3 m spatial resolution) from 3 January 2015 with the same avalanche debris and the GPS track imposed. (c) Oblique field photograph of the avalanche debris with the interpretation of the outline visualized by a pink line Comparison With Manual Identification Statistics of the Detected Avalanche Pixels In sections 4.1 and 4.2 we have shown results of the automatic detection algorithm when applied to two case study areas using the Sentinel-1A backscatter difference images constructed for 6 January 2015 following the New Year avalanche cycle. By comparing the automatically detected avalanche debris with the manually interpreted avalanche pixels we have seen that generally, the automatic detection algorithm succeeds in detecting the majority of the medium to larger sized avalanche debris. However, there remains a significant fraction of manually interpreted avalanche debris that are not captured as well as pixels that have been classified as avalanches debris automatically but not manually. The question we now wish to address in this section is whether we can characterize the pixels that have been correctly and incorrectly classified in order to determine when the automatic detection algorithm is likely to fail or succeed. This may give an indication of the likely measures that may need to be taken in order to improve the classification results. In Figure 9 we show, respectively, for Tamokdalen (left) and Lavangsdalen (right) the histograms of the backscatter difference for all pixels that were correctly classified by the automatic detection algorithm ( correct, green), pixels that were missed ( omission, blue), and pixels that were misclassified as avalanche debris ( commission, red). These histograms are calculated using the detection results presented for 6 January 2015 acquisitions in Figure 4 (Tamokdalen) and Figure 7 (Lavangsdalen). Note that while we have used backscatter difference images for both copolarization and cross-polarization channels in the algorithm, the ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 456

12 Figure 9. (left) Histogram of the automatically detected avalanche pixels for Tamokdalen using the backscatter difference image for 6 January 2015 and separated into three categories: missed pixels, misclassified pixels, and correctly classified pixels. (right) The equivalent histograms for Lavangsdalen. histograms shown in Figure 9 are presented only for copolarization channel (VV). In both cases the histograms are normalized such that the total area under each curve is equal to 1. This has been done to account for the differences in total number of avalanche pixels between the two regions and therefore make the two figures comparable. For the detections in Tamokdalen we see that the distribution of backscatter difference for the correctly detected pixels (i.e., where the automatic detection algorithm detected the same pixels as those identified manually) has similar width compared with the distributions for the missed and incorrect detections but is shifted toward higher backscatter difference values. For all three types of detections (correct, missed, and misclassified) the standard deviation of the backscatter difference values was 2.09, 2.20, and 1.99 db, respectively. The tendency for the distribution peak to lie at higher backscatter difference values is also the case for the correct detections in Lavangsdalen, but here the distribution has a double-peaked shape with maxima at approximately 7.00 and 9.00 db. For both areas, the peak of the distribution for the misclassified pixels lies at a slightly lower backscatter difference than those pixels that were missed. The standard deviation of the backscatter difference values for the three types of detections at Lavangsdalen were slightly higher (2.35, 2.56, and 2.62 db, respectively, for correct detections, missed pixels, and misclassified pixels, respectively), as reflected by the slightly broader curves in Figure 9 (right). The mean backscatter difference values corresponding to the distributions shown in Figure 9 are given in Table 2. Even though these histograms represent only the results for automatic detection of avalanche debris for a small subarea of a single Sentinel-1A scene, the mean values of 7.91 db (Lavangsdalen) and 7.09 db (Tamokdalen) in backscatter change indicate that the threshold of 6.00 db is appropriate for achieving correct detections under the geophysical conditions associated with this acquisition, but we note that in reality, the backscatter difference values within these correctly detected avalanche debris can range from 0 to 13 db. In order to capture more of the avalanche debris pixels that were missed we could of course reduce the backscatter threshold that is used to search for potential avalanche debris. However, given that the peak of the incorrectly classified pixels is also Table 2. Mean Backscatter Difference in Decibels, Calculated for (1) All Correctly Detected Pixels, (2) All Pixels Missed by the Automatic Detection Algorithm, and (3) All Extra Pixels Classified as Avalanche Debris by the Automatic Detection Algorithm but Not Manually a Mean Backscatter Difference (db) Tamokdalen Lavangsdalen Correct pixels 7.09 (3657) 7.91 (2284) Missed pixels 4.90 (2212) 4.89 (1433) Misclassified pixels 4.01 (3166) 3.28 (2821) a The figures in brackets indicate the number of pixels contributing to the calculation of the mean. located at a lower backscatter difference, this indicates that by lowering the threshold we may also increase the probability of misclassification. This threshold sensitivity is addressed in section Sensitivity to Threshold In this section we address the issue of how accurate the algorithm is in terms of correctly classifying pixels ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 457

13 Table 3. A 2 2 Contingency Table Calculated for Tamokdalen Using the Sentinel-1A Image Acquired on 6 January 2015, When a Threshold of 5, 6, and 7 db Was Used in the Automatic Avalanche Debris Detection Algorithm a Auto detection Avalanche Debris Not Avalanche Debris Total Manual Interpretation (5 db) Avalanche x = 3,962 z = 5,121 9,083 Not avalanche y = 1,907 w = 1,139,654 1,141,561 Total 5,869 1,144,775 1,150,644 Manual Interpretation (6 db) Avalanche x = 3,657 z = 3,166 6,823 Not avalanche y = 2,212 w = 1,141,609 1,143,821 Total 5,869 1,144,775 1,150,644 Manual Interpretation (7 db) Avalanche x = 3,218 z = 1,223 4,441 Not avalanche y = 2,651 w = 1,143,552 1,146,203 Total 5,869 1,144,775 1,150,644 a x is the number of pixels correctly detected by the automatic detection algorithm, y is the number of missed pixels, z is the number of false detections (misclassified), and w is the number of correctly classified nonavalanche pixels. as avalanche debris. The use of contingency tables and measures of skill have been documented and applied in the context of forecast verification [Doswell et al., 1990]. Contingency tables are essentially a representation of the number of correctly forecasted events (or in our case, detected avalanche debris pixels) compared with the observed number of events (i.e., manually identified debris). From these quantities, a number of parameters may be calculated. Of interest in this study are the probability of detection (POD), false alarm ratio (FAR), and frequency of misses (FOM) from which further measures of skill may be calculated. The true skill statistic (TSS) is essentially the difference between the probability of correct detections and probability of unexpected nonevents. In the case where the probability of nonevents is high, TSS converges to the probability of correct detections [Doswell et al., 1990]. On the other hand, the Heidke Skill Score takes into account correct detections that were not due to chance and is thus more appropriate in the case where rare events can occur. Haklander and Van Delden [2003] have taken this a step further and combined the best of the two skill scores to formulate a normalized skill score (NSS). NSS is defined as NSSðσÞ ¼ 1 sðσþ þ TSSðσÞ (1) 2 S max TSS max where σ is a variable threshold (in this case the backscatter change threshold) and S max and TSS max are the maximum values of S and TSS. We consider all three skill scores in our assessment of the algorithm s performance when different backscatter change thresholds are applied and how the results may be used in future determination of the optimal threshold for the algorithm. The reader is referred to both Doswell et al. [1990] and Haklander and Van Delden [2003] for a more detailed definition of the parameter calculations. Since we do not have ground truth data for all avalanche debris in the Tamokdalen, the manually identified avalanche debris pixels are taken as a proxy for the observed debris for the purpose of the calculations. We use the manually identified avalanche debris pixels for Tamokdalen corresponding to the Sentinel-1A image acquired on 6 January Table 3 shows the contingency tables calculated for the automatically detected and manually identified debris pixels when thresholds of 5, 6, and 7 db are applied to the algorithm, respectively. In Table 4 we summarize the parameters POD, FAR, FOM, TSS, S, and NSS for the three different thresholds. It is clear that the threshold applied, which determines whether a particular subwindow of the Sentinel image contains potential avalanche debris, can have a significant influence on the classification outcome. A lower threshold enables the algorithm to achieve a higher POD (fraction of correctly classified pixels) and lower FOM (fewer avalanche debris pixels missed) but on the other hand is associated with a higher FAR (fraction of false positive pixels). We see that as the threshold is increased, the automatic detection algorithm misses a greater fraction of the manually identified avalanche pixels, but there is also a reduction in FAR. We note, as shown by Doswell et al. [1990] that the TSS and POD are almost identical for all three thresholds since the correct detection of nonavalanche debris pixels dominates and thus TSS decreases with increasing threshold. This would tend to suggest that the TSS is in this case not a very useful measure of skill of the detection algorithm since the incorrect classifications are more or less disregarded due to the large number of correct nonavalanche pixel detections. Both the Heidke Skill Score and Normalized Skill Score on the other hand exhibit an increasing trend with increased backscatter change threshold, implying that the results ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 458

14 Table 4. Summary of the Parameters Used to Evaluate the Performance of the Automatic Avalanche Debris Detection Algorithm Under Applied Thresholds of 5, 6 and 7 db Threshold POD FAR FOM TSS S NSS 5.0 db db db associated with greater probability of missed detections is associated with the best skill score. From an operational point of view it is important to consider choosing a threshold where an optimal balance between the probability of correct detections, false detections, and missed avalanche debris pixels is achieved. It may be argued that it would be preferential to choose a lower threshold and overdetect avalanche debris pixels than select a higher threshold where a greater number of avalanche debris pixels are missed. On the other hand, if this result is likely to have a strong influence on decisions, for example, whether to evacuate a populated area because of apparent high avalanche activity, then the result could involve unnecessary costs in order to execute an evacuation operation. 5. Discussion 5.1. Misclassification The automatic avalanche debris detection algorithm presented in this study is dependent on both change detection and object classification techniques applied to SAR backscatter. In the first stage it is the choice of reference image, which largely determines what features are visible when the backscatter difference image is constructed and processed at the classification stage. The observed SAR backscatter varies according to the snow cover, which is also dependent on weather conditions. As noted earlier, the snow condition in the reference image should ideally be very similar to the snow condition of the undisturbed snow in the avalanche debris image such that the backscatter difference image highlights those areas where avalanche debris has resulted in an increase in backscatter. Some nonavalanche areas exhibit increased backscatter between two points in time and appear to have similar properties as avalanche debris, leading to incorrect classification. This can and does occur if, for example, the reference image contains wet snow (due to warmer weather at the time when the reference image was acquired) and the avalanche debris image contains dry snow, as a result of colder conditions when the avalanche image was acquired. A backscatter difference image computed from a reference image with wet snow and an avalanche image containing dry snow will lead to image pixels which have increased backscatter, since the radar backscatter from wet snow is usually lower than from dry snow. However, the increase in backscatter does not result because of increased surface roughness as would be expected from avalanche debris but as a result of the change in the snow scattering contributions. Reference images with wet snow are hence challenging for the present algorithm and should be treated differently or at least flagged. Low-lying vegetated areas may also exhibit backscatter change similar to that of true avalanche debris thereby leading to wrong classification and higher commission errors. Misclassification of snow-covered trees and bushes has been identified as a problem by Bühler et al. [2009] when they applied an object classification procedure to optical remote sensing data. However, we have deliberately not included forest and vegetation in the masking process at this stage in order to avoid incorrectly excluding vegetated areas that have been previously observed in runout zones. Only by building up a thorough database and parameterizing avalanche occurrence will it be possible to construct a more accurate mask which identifies the most avalanche-prone zones. This will undoubtedly improve the algorithm by reducing its dependence on thresholding. Also, masking out all zones where the slope is greater than 35 is somewhat simplistic and we cannot rule out the possibility that debris can be found outside of this slope threshold. Even avalanche debris on icecovered lakes (belonging to the water mask) can occur. In order to refine the slope mask we would ideally need to obtain more precisely mapped avalanche release and runout zones. This could be done initially through manual detection of the paths or computationally using paths generated by numerical simulation models [Bartelt et al., 2012; Bühler et al., 2013] Algorithm Variables Another parameter in the algorithm that may affect the classification accuracy is the window size. When the window contains only two distinct populations, i.e., avalanche pixels and nonavalanche debris pixels, the likelihood of a correct classification is greater. The classification algorithm requires a certain minimum ECKERSTORFER ET AL. AUTOMATED AVALANCHE DETECTION 459