MSG/SEVIRI AND METOP/AVHRR SNOW EXTENT PRODUCTS IN H-SAF Niilo Siljamo, Otto Hyvärinen Finnish Meteorological Institute, Erik Palménin aukio 1, Helsinki, Finland Abstract Weather and meteorological processes are affected by the varying snow cover and snow properties. Snow areal extent is essential information for weather forecasting and nowcasting. Successful snow detection is possible using data from either polar orbiting satellites or geostationary satellites (e.g. H- SAF snow cover products based on Metop/AVHRR and MSG/SEVIRI data). Polar orbiting satellites offer the best coverage and resolution in the polar regions, but only a limited temporal resolution. Geostationary satellites have poor resolution in polar regions, but better temporal resolution. The H-SAF snow cover products presented in this work were developed originally in the LSA SAF, but since 2012 they have been part of the H-SAF. In this work, we describe the status and some validation results of the current H-SAF MSG/SEVIRI and Metop/AVHRR snow detection products. The snow extent products are traditionally validated using surface observations (snow depth) from synoptic weather stations. Another source of validation data are snow cover measuring campaigns, such as SEER, but these can not be used for continuous validation. A novel non-automatic validation method for snow cover products using photographs from social media services has been developed and some results will be presented. CURRENT STATUS OF THE FORMER LSA SAF SNOW COVER PRODUCTS The MSG/SEVIRI snow cover product is operational and available from the LSA SAF web site (http://landsaf.ipma.pt). The Metop/AVHRR PDU based snow cover product is currently generated for internal use in the LSA SAF system. These PDU based SC products will be used to generate the global Metop/AVHRR snow cover product in the future. ALGORITHM AND PRODUCT STATUS During the CDOP phase of the LSA SAF, two algorithms for detection of snow extent were developed. The algorithms classify each pixel as snow free, partially snow covered or totally snow covered based on satellite data. For the daily MSG/SEVIRI snow cover product, all snow cover maps which are produced every 15 minutes are combined. The algorithm counts the number of different classifications for each pixel and then determines the final daily classification if there have been reasonable amount of cloud free observations during the day. The snow extent algorithm for Metop/AVHRR is employed to generate PDU based snow extent products. These will be combined to generate a daily snow extent product. The preliminary version of the global product in the Figure 1 shows that the snow in the high latitudes is detected quite well also in North America and Asia.
Figure 1: Global compilation of the PDU based Metop/AVHRR snow cover products during March 25 th, 2012. Green snow free areas cover most of the land. Snow (white) is present in the northern regions and mountains. Red marks the unclassified pixels. The snow covered areas of the Northern Hemisphere are detected accurately. However, there are some thunderstorms (i.e. ice clouds) over the Amazon region in the South America, which are not classified correctly as clouds. VALIDATION Snow cover products should be validated using the surface observations of the snow coverage. Unfortunately such data is not widely available. Synoptic weather stations measure snow depth which is not directly related to snow coverage especially during the melting season and in the forest areas. The reporting practices of the weather stations create additional challenges because the stations do not report snow depth when there is no snow at the station. This makes is almost impossible to know whether the station is not measuring snow or there is no snow at the station. The lack of direct and operational observations of the snow coverage forces us to consider alternative data sources. Several possible data sources exist and have been considered. Practical but rather limited method is to compare the product to other satellite snow cover products. On the other hand, surface observations of the snow coverage can be obtained from field campaigns such as Snow Edge and Extent Retrieval (SEER) campaign in which snow coverage data was collected from a moving car. Other possibilities are unmanned aerial vehicles (UAV) which can be used to take images of the snow cover. However, the analysis of these images is laborious. The most novel approach is to use social media photographs as a source of snow cover data. Other satellite product We compare our product to other satellite snow cover products. This is reasonably reliable method if the baseline snow cover product is documented and validated. For example Interactive Multisensor Snow and Ice Mapping System (IMS) has a good reputation. We have used it as a base line for MSG/SEVIRI SC product validation (Siljamo & Hyvärinen, 2011). The comparison results of the MSG/SEVIRI SC and IMS are presented as time series in Figure 2. The maximum extent of snow is in February; the snow starts to melt in mid-march and has melted almost
completely by the end of May. The remaining snow is found on mountaintops and glaciers, which are mostly outside the scope of these SAF snow cover products. New snow starts to accumulate in November. The surface area that is classified varies both because of the varying cloud cover and also with the season, as during the winter the zenith angle can be too high to enable classification for all day in northern regions. In addition, the results should be given less emphasis when the absolute amount of snow is very small in summer, as the results vary greatly owing to even slight differences in snow cover between products. Therefore, days when there are more than 20 times more correct rejections than other classes are depicted with a different colour in the rest of the plots in Figure 2. Figure 2: Amount of cloud-free land pixels (blue dashed line, IMS; red circle, version 1; black circle, version 2), and amount of snow-covered pixels (blue line, IMS; red line, version 1; black line, version 2). (b) Bias, (c) FAR, (d)h, (e) PC, and (f) HSS for version 1 (red circle) and version 2 (black circle) when compared with the IMS product. When the correct rejections exceed the other classes by more than 20 times, version 1 is shown in pink and version 2 in gray crosses. Vertical dotted lines show the transition from version 1 to version 2.02 (red), version 2.05 (gray), and version 2.10 (black). Curves show the two-month moving average of the data. From Siljamo and Hyvärinen (2011).
Campaign data To collect some snow coverage data and to test feasibility of the collecting coverage data from a moving vehicle the Snow Extent and Edge Retrieval (SEER) campaign was arranged in April 2011. By driving from snow covered area to snow free area and back we were able to collect snow coverage data. The snow coverage near the road was estimated and saved with location, photographs and some auxiliary measurements. The vehicle and the driving route during the campaign are presented in the Figure 3. During the campaign snow was melting fast and the edge of the snow covered area was moving daily. Reasonably full snow cover could be found from eastern and northern parts of Finland. Figure 3: The measuring vehicle used during the SEER campaign on the left. The SEER route is shown on the right. Green markers show the start and stop of each measuring day. (Map Google 2011) Although a rough estimate of the percentage of the snow covered area can be estimated from a moving car, there are several factors which must be considered when using the data. There is no natural undisturbed snow near the roads, because snow removed from the road can change the properties of the snow along the roads. Snow tends to melt fasters near the road, also. The snow coverage should be estimated by observing the snow cover at least 20-50 meters from the road to be able to avoid disturbed snow. Other factors to be taken in to account are land cover type, lakes and other water bodies which can be frozen or open and topography. The roads are not usually built on rough terrain or wetlands which also change the snow cover properties. The observed snow coverage was averaged for each MSG/SEVIRI and Metop/AVHRR pixel. Figure 4 shows the results of a preliminary comparison of the observed snow coverage and snow product classification for each pixel along the SEER route. The snow cover products resemble the observed snow coverage well although the effects of topography, lakes and land cover class are not yet considered.
Figure 4: The comparison of the SEER snow coverage data to MSG/SEVIRI snow cover (on the left) and Metop/AVHRR snow cover (on the right). Vertical stripes show the surface classification on the satellite product in each pixel along the SEER route. Dark green shows snow free pixels, light green partially snow covered pixels and grey snow covered pixels. White stripes are unclassified (usually cloudy) pixels. Black lines show the sliding average of the snow coverage estimates along the SEER route. Unmanned aerial vehicles (UAVs) Weather stations and other point like observations of the snow cover, such as SEER data, provide usually only a rather limited number of observations of the snow cover. These observations might not be representative especially in the forests and in the areas further away from the observation point. UAV based images provide better areal coverage and better estimates of the snow covered area, but the data analysis is more labour intensive than point observations. During the spring 2013 a small scale feasibility campaign was organized with Novia University of Applied Sciences (Tammisaari, Finland) and Itä-Lapin ammattiopisto (vocational college, Kemijärvi, Finland). UAVs were used to take photographs in three days during the melting season. The analysis of the data is not finished, but the advancing snow melt is quite clearly presented in the photos. Trees and other obstacles change the melting patterns and also create more challenging targets for the snow detection. In the Figure 5 we show an example of the analysis. Simple brightness based thresholding method does not detect most of the snow in the forest when compared to analysis made by hand. The analysis is most challenging in the forest where the trees and the shadows make it quite difficult to see whether the surface is snow covered or not.
Figure 5: UAV photo (top left) and analysis of the snow cover in April 20, 2013 near Fiskars, Finland. Snow cover has been melting rapidly. Open field are almost snow free, but parts of the forest are still snow covered. Top right corner shows a simple thresholding based estimate of the snow covered area. This method suggests that 7.4% of the surface is snow covered. Bottom left shows the hand made analysis on top of the original image and bottom right shows only the analysis. This method produces 41.5% snow coverage. Hand made analysis and simple automatic analysis differ especially in the dense forests where the trees and shadows of the trees make the analysis challenging. Photo by George Rybakov/Novia. Social media A novel method to estimate the surface status is to use images presented in social media such as Facebook, Flickr, Instagram and other services which allow sharing of photographs taken by mobile phones or different cameras. Quite often these photos are geotagged and time stamped so that it is possible to compare the surface status obtained from the satellite product to the analyzed surface status in the images. We chose to use Flickr as the data source for our study. To test the feasibility of the method we retrieved a photoset of about 3000 images. To be able to determine the snow status we used a set of simple rules to classify each image either snow covered, snow free or irrelevant. The images had to be taken outside (and not for example trough the window), the surface near the photographer must be visible and we must see at least some of the surroundings to be able to estimate the probable surface status. For these reasons most of the images can not be used for analysis. Although all of the images had geographical coordinates we found out that in the future analysis we must select only images
where the location is retrieved by a GPS receiver in the camera and the latest location fix is relatively new (less than 5 minutes). First preliminary results of the validation using Flickr images are presented in the Figure 6. Pixels of IMS and LSA SAF images (projected to the same projection) were tabulated to a 2 x 2 contingency table. In Figure 6, the size of rectangles represents the number of pixels in the contingency table. Then photos that were taken inside pixels from each four segments of the contingency table were retrieved from Flickr. These photos were subjective classified to no snow, snow and not relevant classes. The fraction of photos with snow was taken to represent the fraction of snow-covered pixels in each segment of the contingency table. In Figure 6, the four segments are divided by this fraction to snow and snow-free regions. Therefore, for example, we estimate that when IMS detects snow but LSA SAF detects no snow, in more than half cases there are snow in those pixels. In mountains, when both satellite products agree, there are more cases of snow in pixels classified as snow-free. a) Flat lands b) Mountains Figure 6: Pixels of IMS and LSA SAF images (projected to the same projection) were tabulated to a 2 x 2 contingency table. The size of rectangles represents the number of pixels in the contingency table. Then photos that were taken inside pixels from each four segments of the contingency table were retrieved from Flickr. These photos were subjective classified to no snow (green), snow (pink) and not relevant (not presented in the figure) classes. The fraction of photos with snow was taken to represent the fraction of snow-covered pixels in each segment of the contingency table. The four segments are divided by this fraction to snow and snow-free regions. Therefore, for example, we estimate that when IMS detects snow but LSA SAF detects no snow, in more than half cases there are snow in those pixels. Results for the flat areas are presented on panel a) on the left. In mountains (panel b) on the right), when both satellite products agree, there are more cases of snow in pixels classified as snow-free. SUMMARY The former LSA SAF snow cover products have been part of HSAF since the March 2012. Currently operational product will be generated in the LSA SAF system. Current versions of the MSG/SEVIRI global daily snow cover product and the Metop/AVHRR PDU based snow cover product show good agreement with observations. Daily global version of the Metop/AVHRR snow cover product is still under development, but preliminary results are very promising. Traditional surface observations and other satellite snow cover products can be used for product validation. However, preliminary results show that new methods such as UAV imagery and social media photos might provide valuable new data sources for product validation.
REFERENCES Siljamo, N. and Hyvärinen, O., (2011) New Geostationary Satellite-Based Snow-Cover Algorithm. J. Appl. Meteorol. Climatol., 50, June 2011, 1275 1290, doi: 10.1175/2010JAMC2568. ACKNOWLEDGEMENTS This work is financially supported by the EUMETSAT.