SNOW COVER MAPPING USING METOP/AVHRR DATA

Transcription

1 SNOW COVER MAPPING USING METOP/AVHRR DATA Niilo Siljamo, Markku Suomalainen, Otto Hyvärinen Finnish Meteorological Institute, Erik Palménin Aukio 1, FI Helsinki, Finland Abstract LSA SAF snow cover product for MSG/SEVIRI has been available some years. The methods used for the development of the MSG/SEVIRI algorithm have been modified and applied to the development of the LSA SAF MetOP/AVHRR snow cover algorithm. None of the threshold rules used in the SEVIRI algorithm could be directly used in the AVHRR algorithm, but the the development methods worked out well. Thus we have built for AVHRR an algorithm, which is more straightforward than the SEVIRI version, but still works surprisingly well. INTRODUCTION Weather and meteorological processes are affected by the varying snow cover. Snow areal extent is essential information for weather forecasting and nowcasting. Successful snow detection is possible using data from geostationary satellites (e.g. LSA SAF snow cover product based on MSG/SEVIRI data, Siljamo and Hyvärinen (2010)). However, even though these products have superb temporal resolution and thus a couple of hours of cloudless weather per day is enough to determine the snow cover, their limited spatial resolution in the polar regions is a problem. The answer for this problem is polar orbiting satellites. When they fly over the polar regions, we achieve a better spatial resolution compared to surveying from the geostationary orbit. The MetOP satellite is equipped with the AVHRR instrument, which older than SEVIRI and inferior to it when considering the amount of channels and the measurable wavelengths. The list of AVHRR channels and their wavelengths are presented in Table 1. With the polar regions, our main interests lay in measuring the snow cover over a longer timespan, which could be helpful when researching the global warming. The AVHRR instruments have been in use since 1970 s and thus provide us with excellent data set over a long period, if a sufficient algorithm can be constructed. Channel Central Availability wavelength (nm) h h 3a 1161 Day (selectable) 3b 3740 Night ( selectable) h h Table 1: List of AVHRR channels. Channels 3a and 3b cannot be used simultaneously ALGORITHM AVHRR instruments measure the reflectance of the Earth on the 6 relatively wide spectral bands described in Table 1, while the SEVIRI instrument has 12 channels. Some spectral bands which were used in the SEVIRI algorithm to detect snow are not available, so completely new rules are needed

2 This first preliminary version of the MetOP/AVHRR algorithm used for snow detection consists of only two main rules, one for snow and one for snow free ground. They are based on the 2D scatter plots in Figures 1 3. These plots consist of over ~ pixels which are classified subjectively to snow covered, snow free and several cloud covered classes. The idea is derived from a similar algorithm for the MSG/SEVIRI instrument, but because of the differences between the channels in AVHRR and SEVIRI, only the satellite- and sun angle rules could be directly imported from the SEVIRI algorithm. The main rules for snow detection are the following: Land if TB5 > OR C2/C1 > Snow if C2/C3 > 90 OR C2/C3 > 44 AND TB4 > AND TB4 < 269.7, where TB is the brightness temperature and C the radiance on a specified channel. These rules are based on Figures 1 and 2 Figure 3 looks promising regarding some future ideas of the development of the algorithm. There are also the rules for satellite and sun angle. They tell whether the pixel is too dark to be recognizable at all. Figure 1: Brightness temperature on channel 5 and the ratio of radiances on channels 2 and 1

3 Figure 2: Brightness temperature on channel 4 and the ratio of radiances on channels 2 and 3 Figure 3: Natural logarithms of radiances on channels 2 and 3 SNOW COVER MAPS The snow cover maps produced by the algorithm, and the respective MetOP/AVHRR RGB images used for subjective snow cover analysis at the Finnish Meteorological Institute, are presented in Figures 3 7. The RGB maps are combinations of AVHRR channels 3A, 2 and 1. Snow cover can be seen as cyan shades in the RGB images. Clouds are also cyan (ice clouds) or white (water clouds). In

4 the snow cover maps snow is white, snow free land is green and red marks cloudy, dark or otherwise unclassified areas. Figure 4: March 09, 2010, 0834UTC, RGB map Figure 5: March 09, 2010, 0834UTC, snow cover map March , presented in Figures 4 and 5, was a typical spring day. There was some snow in Central Europe, which the algorithm seems to detect quite well. Snow free land is also classified correctly. Cloudy areas are unclassified.

5 Figure 6: February 02, 2010, 0858UTC, RGB map. Figure 7: February 02, 2010, 0858UTC, snow cover map The second example from 02 February 2010, presented in Figures 6 and 7, shows that the conditions for snow detection are much more difficult during the mid-winter. There is not much light and the weather is often cloudy. The surface is difficult or impossible to see from the satellite through the clouds and even in cloud free regions darkness or low sun elevation angle can prevent the use of the algorithm.

6 Figure 8: Details, March 09, 2010, 0834UTC Figure 9: Details, March 09, 2010, 0834UTC The images in Figures 8 and 9 show the classification result of the preliminary version of the snow cover algorithm. Coloured dots present the classification of each pixel in the RGB image. There is also a line showing an approximate edge of the snow cover detected by the algorithm. Although there are still some differences and misclassifications, the algorithm seems to produce realistic results. When the algorithm was tested for the training data set, it classified correctly about 91% of the pixels, which we consider to be above expectations for a first development version of the algorithm.

7 WHAT HAPPENS NEXT The next phase of the development will concentrate on finding the reasons for misclassifications. Especially in areas near the edge of the snow field a special class of partial or probable snow could be used to enhance the product. For the future, we have two new ideas to develop the algorithm. The first is to use Support Vector Machines (SVM) with transformation; the concept is to use a kernel function to project the data into more dimensions, where the classified pixels could be separated linearly or near linearly using a slack variable, which gives penalty for misclassified pixels with the SVM. The theory of SVM is excellent and proven to work with smaller datasets, but the first attempts with the classified snow cover data have showed that the computational cost might be too high. Also, more research needs to be done regarding the choice of the kernel, since the number of support vectors revealed severe overfitting. The second idea is to fit a curve into the classified data sets in a 2D scatter plot, and measure the distance from each data point to each of the three curves snow, snow free and cloud. The ratio of the distances would give us probabilities of a pixel belonging to a certain class. CONCLUSIONS Preliminary results with the AVHRR are better than expected after all, it is rather limited compared to SEVIRI in terms of channels. The AVHRR algorithm hasn t obviously reached as good results as the SEVIRI algorithm developed earlier in FMI, but the start is really promising, especially because there are so far only 5 rules alltogether in the AVHRR algorithm, compared to the ~20 rules in the SEVIRI one. We are also hoping to make either of the ideas mentioned previously in the what happens next -chapter work, which would move us out from the concept of rules and towards statistical classification using percentages instead of threshold values. REFERENCES Siljamo, N. and Hyvärinen, O., (2010) New geostationary satellite-based snow cover algorithm. Journal of Applied Meteorology and Climatology, submitted