Embedded Files
Active radar sensors on board of the Sentinel-1 satellites use microwave radiation to illuminate Earth's surface and are thus independent of weather and light conditions. The Sentinel-1 satellites deliver free SAR data in a consistent manner, enabling near-real time monitoring of medium sized avalanches almost anywhere on Earth.
Active radar sensors on board of the Sentinel-1 satellites use microwave radiation to illuminate Earth's surface and are thus independent of weather and light conditions. The Sentinel-1 satellites deliver free SAR data in a consistent manner, enabling near-real time monitoring of medium sized avalanches almost anywhere on Earth.
Radar sensor parameters
Radar sensor parameters
The Sentinel-1 constellation consists of two identical satellites (S1A & S1B) that orbit Earth 180 deg apart. The satellites fly in a sun-synchronous, near-polar orbit with the radar instruments illuminating the surface perpendicular to the flight direction. The satellites fly towards and from the North Pole in an ascending and descending orbit respectively. At either given orbit, particular slope aspects are illuminated, while little to no information is gained from the opposite aspect.
Avalanche detection using microwave radar
Avalanche detection using microwave radar
In case of an avalanche releasing, relatively more energy is scattered back to the radar sensor from avalanche debris than from undisturbed snow. The higher backscatter stems from the relatively high surface roughness that avalanche debris exhibits.
Avalanche detection is based on a temporal change detection method, where relative backscatter change between two radar images of similar satellite orbit is visualised in a change detection image. Revisit frequency of the Sentinel-1 constellation is thus of importance. The temporal baseline of change detection images worldwide is a minimum of 6 days and a maximum of 12 days. Only from increased coverage frequency using multiple orbits, the uncertainty in release timing can be reduced to a minimum of 1 day and a maximum of 3 days. This, however, requires the availability of both ascending and descending orbit data.
Sentinel-1 data in interferometric wideswath mode (IW) has a ground range resolution of 20 x 5 m which we usually resample to 20 m pixel spacing. A spatial resolution of 20 m allows for detection of medium sized avalanches with a typical path length of 100 m. Such avalanches can bury and kill a person.
Temporal change detection
Temporal change detection
The relative backscatter increase from avalanche debris in change detection images is used for detection. A clear backscatter threshold between avalanche debris and undisturbed surrounding snow does not exist, however. The reason is the dynamic nature of the snowpack, which constantly changes its dielectric properties experienced by the radar wave. Moreover, avalanche debris and surrounding snow are comprised of the same material.
We construct change detection (c) images by combining single backscatter images from a reference image (a) and an activity image (b) that depicts avalanche activity.
In an ideal case, a change detection image displays a transition from dry to wet snow conditions, which results in a relative decrease in backscatter everywhere in the landscape except for avalanche debris which become easily detectable.
Automatic avalanche debris detection
Automatic avalanche debris detection
The algorithm starts by (a) specifying and geocoding SAR data, associated DEM and masking files (b), remapping to the area of interest. The geocoded SAR images are divided into tiles and a difference image is constructed between activity image (with avalanche activity) and reference image (c). A difference of Gaussian filter is applied to the difference image and the mean and standard deviation of pixels in the filtered image are used to automatically calculate a threshold for detection of avalanche debris. Then the pixels are ranked (d) based on their backscatter values and divided into six equal classes. The class means are defined and used to segment each tile in order to determine which class mean is closest to the backscatter value of each pixel. The class change is then calculated for the segmented tiles and the mean and standard deviation are used to obtain a threshold (currently at 4.0 dB) (e). In a final step (f), detected pixels common to both edge detection and class change are kept as avalanche pixels (min size 15 px).
Algorithm performance
Algorithm performance
The avalanche detection algorithm has an average probability of detection (POD) of about 70 % compared to manual interpretations. The performance is, however, largely affected by the temporal change detection and thus POD’s can range between 100% and 50 %. The false positive rate (detection of features that are not avalanches) is typically 30% or greater. This means that the total number of detections is at least the same as the number identified manually, but not all of them are true avalanches. A qualitative analysis of detected avalanche types yields that avalanches with high surface roughness such as wet snow avalanches have a higher POD than for example soft slab avalanches or loose snow avalanches.
Automatic processing chain
Automatic processing chain
The avalanche detection algorithm is built into an automatic processing chain for use of avalanche activity detection in operational avalanche forecasting. For a single Sentinel-1 scene, it takes roughly 50 min from initial SAR processing to export of avalanche activity into a geodatabase. The processing chain is thus capable of delivering avalanche detections in near-real time, consistently throughout a winter.
Workflow from avalanche activity, its raster-based automatic detection using temporal change detection, to vectorisation and storage in a geodatabase with metainformation.
Report abuse