Remote Sensing and Hydrology 2000 (Proceedings of a symposium held at Santa Fe, New Mexico, USA, April 2000). IAHS Publ. no. 267, 2001. 333 Estimation of snow cover over large mountainous areas using Radarsat ScanSAR HAROLD HAEFNER, DAVID SMALL, STEFAN BIEGGER, HILKO HOFFMANN & DANIEL NÙESCH Remote Sensing Laboratories, Department of Geography, University of Zurich, Winterthurerslrasse 190, CH-8057 Zurich, Switzerland e-mail: haefner@rsl.geo.unizh.ch Abstract As part of our general objective to develop and operationalize methods for snow monitoring in high mountain terrain, the possibilities of snow cover mapping over large mountainous areas e.g. almost the entire Swiss Alps and Pre-Alps using Radarsat ScanSAR, and NOAA-AVHRR for comparison, are evaluated. The generation of snow masks in AVHRR scenes is achieved using a simple threshold-based method. A change detection approach is employed to compare the two SAR-scenes. It is expected that the most important differences will result from vast changes in the snow cover. Since large parts of the Swiss Plateau were also snow covered in the winter scene, other changes in land cover are expected to have only minor influence. Hence, it will be possible to assess the total snow cover by combining the information gained from the SAR ratios with the relief information from the DEM. Results from the two sensors are compared with each other, and with ground-based meteorological observations. Verification using a finer temporal resolution is desirable. Key words change detection; monitoring of large mountainous areas; NOAA-AVHRR; Radarsat ScanSAR; snow cover estimation; Swiss Alps INTRODUCTION A very decisive factor for sustainable development of mountainous areas and for integrated, careful management of their fragile ecosystems, is the seasonal variation of the snow cover. Therefore, precise, detailed, long-term systems for monitoring spatiotemporal snow cover variations, and the compilation of comparable, continuous snow pack inventories at scales from local watersheds to whole mountain systems and even global overviews, have to be established. In addition to areal extent and magnitude, snow water equivalent and snow pack wetness are the most important parameters to be determined. To reach this goal, different satellite remote sensing systems in combination with GIS and DEM, have to be used effectively. Areal extent and areal water equivalent can be measured with optical sensors (Ehrler et al, 1997), but continuous mapping in middle and high latitude climates can only be achieved with microwave data (Haefner et al, 1998). Combining the advantages of EO and SAR systems by using data fusion technologies will improve the results. An approach has been developed for continuous, precise monitoring of regional snow cover with ERS-1/2 SAR data for an entire accumulation and melting period. This method, the "multitemporal optimal resolution approach" (MORA), has been tested on the micro- and mesoscale level (Piesbergen et al, 1998). Based on these experiences, an extension to larger mountainous areas to simultaneously monitor the
334 Harold Haefner et al. snow pack of vast, complex watersheds, or of several connected watersheds, with SAR data, and its correlation with EO data, is currently under investigation. This paper presents the first results of this work from the Swiss Alps, discussing the possibilities and problems of small-scale mapping of the wet snow cover. STUDY AREA, SATELLITE IMAGERY AND AUXILIARY DATA Two Radarsat ScanSAR acquisitions for 22 February 1999 and 16 July 1999 were available, covering the central and eastern parts of the Swiss Alps, as well as the central parts of the Pre-Alps, the Swiss Plateau and the Jura Mountains. Figure 1 shows an example of this imagery, terrain-geocoded to Swiss map coordinates. The single scene covers an area of approximately 300 km 2 with a resolution of about 50 m. 3501 ' Switzerland - Terrain-Geocoded ScanSAR (W1+W2) 1 1 m 1 i ' 1 1 1 - ^ - ^ 600 700 800 Fig. 1 Terrain-geocoded Radarsat-1 ScanSAR narrow, 22 February 1999. Canadian Space Agency, 1999. Two cloud-free NOAA-AVHRR images acquired as close to the SAR image as possible were selected for comparative studies. For the winter scene, a time difference of four days had to be accepted. For the summer scene, a useful NOAA-image acquired just one day later (17 July) exists (see Fig. 3). A DEM with a grid spacing of 50 m was available for all of Switzerland. Outside of that area, the GTOPO30 global elevation model (1 km spacing) was used. For the evaluation of the results, the records of 72 meteorological stations regularly distributed over the test area were used to get a clear understanding of the temperature and precipitation developments before, during and after the overflight. However, it should be noted that few of the stations are located at higher altitudes, i.e. above 1500 m (Fig. 2).
Estimation of snow cover over large mountainous areas using Radarsat ScanSAR 335 Air Temperature, February 22,7.00 am (+), July 16,8.00 am (o) : Freezing Level Jury 16, 1999 : 3100 m.a.sx O o : O o^<s><? Freezing Level February 22, 1999 Ï134 rri'.a.s.i. CPS o -15-10 0 5 10 Temperature [deg] Fig. 2 Temperature profile at meteorological stations (SMA ANETZ) during Radarsat acquisitions. SNOW COVER ESTIMATION NOAA-AVHRR Many different methods have been developed for the generation of the snow and cloud mask in AVHRR scenes (e.g. Derrien et al, 1993; Simpson & Gobat, 1995). One can organise these roughly into: (a) statistical, (b) threshold, and (c) pattern recognition based methods. Newer work describes the use of neural networks for generating cloud masks (Yhann & Simpson, 1995). For the results presented here, a simplified threshold-based method adapted from Derrien et al. (1993) was used. One advantage of the method is that it has been in operational use since 1993; its strengths and weaknesses are generally well known. The principle is based on pixel-wise query of the individual channels or channelcombinations, and testing the values against the defined thresholds. The threshold values were set interactively for each scene: if (ch 1 > ch 1 threshold) and ((ch 3 - ch 4) < (ch 3 - ch 4 threshold)) then snow. The NOAA images and snow masks for the winter and summer scenes are shown in Fig. 3. Quality control was performed via visual inspection. The following difficulties, partially documented in Derrien et al. (1993), were observed in the differentiation between snow and cloud: - In the summer scene in particular there is a risk of mis-classifying stratocumulus clouds as snow. - Visual control is difficult in high mountain areas. Low cumulus clouds above the mountain peaks are visually almost indistinguishable from snow covered areas in such regions. - In the winter scene, the low solar angle, the long shadows and low illumination make the channel 1 test ineffective. - Mixed pixels that include both snow covered and snow free areas are found
500 600 700 800 900 500 600 700 800 900 400 T 300 op 200 2 100 <3* ' mmsam 400 300 m 200 Z 100 0 500 600 700 800 900-100 500 600 700 800 900 350 [ (a) Fig. 3 Geocoded winter (top) and summer (bottom) NOAA images and snow masks. (b) 180 300 250 170 h 160 c: I Z 150 WHSÊÊÊÊIÊÊKBÊBÊ e IS t: o 2 200 150 100 50 lill 600 700 800 130 610 620 630 640 650 660 670 63 i3o r 120 j 110 j 00 J 1 100 O 2 I 90 j so fêêêêêbêêmêmêêmêêêêbsêê H H B B H H H H H I HHMBHBHHHMMHBHMI B B B H H B H H H B M H H H H H B G S HHSHMHBHHHHHHH9I 70 580 590 600 610 620 630 64 Fig. 4 (a) Terrain-geocoded Radarsat ScanSAR February/July backscatter overlay, and (b) close-ups for Bernese Oberland (left) and Valais, Switzerland (right). Canadian Space Agency, 1999.
Estimation of snow cover over large mountainous areas using Radarsat ScanSAR 337 Berner Oberland: Distribution of backscatter differences as a function of altitude 4100-43001- :! + -; - =-? - - + ; \ \ h 39oo-4i oo - ; ; - ; = 4 = \ \- : -j 3700-3900 h : ; r - - = i 1 t : ; h 3500-3700 h ; ; ; ; = ; ; -j 3300-35001- ; ; ; : -4-4 1- - -I 3100-3300j- ; i : -H : f -i 2900-31 oo i- : i :--= ; --; s r- -i 2700-2900H : ; ;-=»=J»- F ;-- tmi + -[ 2500-27001- ' I"*^-! I -l^bfrm«gban--h- - 2300-2500 h > H ; I - I ; - «j a w ; -I 2100-2300 h ; I H=! i* afo- ; H 1900-21 oo h ; + ; : mwii'i'mhmt * ; h i 7oo-i 900 h ++> ' ; : ;mi»imji*+h«h- +; h 1500-1700 H -w-l 4--- I ', ^ lllli irjl + ; -I 1300-1S00F * 7 ', "R= ', ', ', I -I iioo-i3oof -H-j» : «=4=-! -tfyitf ; ; - 900-1100F t*<! ~ t = - - - T -H»!! -I 700-900 - jmmi 7-4= I : ; H 500-700 h I ; I 1 F-gH- ; - -15-10 -5 0 5 10 15 20 25 Backscatter Difference IdB] 1 st Quartile - 3xlQR 3rd Quartile + 3xlQR ; ^ iqr i Outliers 1st Quartile Median 3rd Quartile Outliers Fig. 5 Distribution of RADARSAT backscatter differences (winter-summer) by altitude for the Bernese Oberland, Switzerland. particularly in the summer scene. Since their interpretation is ambiguous, these pixels are not carried forward for further investigation. The threshold values in the snow test were chosen to minimize the number of mixed pixels. - The masking method described above works well only when the test area within the scene is completely cloud free. - Ground fog in the winter scene hinders reliable snow masking. The NOAA snow masks serve in the following sections as an aid to interpreting the Radarsat scenes. The use of AVHRR scenes also offers the advantage of a high repetition rate (multiple overflights per day). This enables capture of the development of the weather situation and snow cover over time for use as a priori knowledge while interpreting the Radarsat scenes. In addition, the brightness temperature from channel 4 provides one indicator for estimating the current snow temperature and snow water equivalent. Radarsat ScanSAR Snow cover estimation has been performed using single beam ERS and Radarsat products in the past (Baghdadi et al, 1997; Guneriussen et al, 1999; Nagler et al,
338 Harold Haefner et al. 1998), but results using ScanSAR imagery have not yet appeared in the literature. ScanSAR processing poses both radiometric and geometric calibration challenges (Srivastavae?a/., 1999). A change detection approach is employed to compare the amplitude differences of the two SAR-scenes, calculating the ratio values. The overlaid images of radar backscatter are shown in Fig. 4(a and b), with close-ups of the Bernese Oberland and the Canton of Valais. Note the reddish areas surrounding mountain peaks in the Bernese Oberland and Valais areas. The backscatter in these areas was lower in the summer (green) than the winter (red), probably due to wet snow cover at those altitudes. The winter snow cover at the same area was most likely dry. Figure 5 shows the altitude-dependency of the backscatter difference between the winter and summer acquisitions. Note the trend towards a positive difference at high altitudes (dampened backscatter from wet snow in the summer vs dry snow in the winter). Likewise, a trend towards a negative difference at lower altitudes (dampened backscatter from wet snow in the winter versus no snow in the summer) is apparent. CONCLUSIONS AND OUTLOOK We have presented a new result showing that it may be possible to track wet snow cover over large mountainous areas, though this will need to be verified using more imagery with a finer temporal distribution. REFERENCES Baghdadi, N., Gauthier, Y. & Bernier, M. (1997) Capability of multitemporal ERS-1 SAR data for wet-snow mapping. Remote Sens. Environ. 60, 174-186. Derrien, M., Farki, B., Harang, L., LeGleau, H., Noyalet, A., Pochic, D. & Sairouni, A. (1993) Automatic cloud detection applied to NOAA-11/AVHRR imagery. Remote Sens. Environ. 46, 246-267. Ehrler, C, Seidel, K. & Martinec, J. (1997) Advance analysis of snowcover based on satellite remote sensing for the assessment of water resources. In: Remote Sensing and Geographical Information Systems for Design and Operation of Water Resources (ed. by M. F. Baumgartner, G. A. Schultz & A. I. Johnson) (Proc. Rabat Symp., April 1997), 93-101. IAHS Publ. no. 242. Guneriussen, T., Johnson, H. & Lauknes, I. (1999) RADARSAT, ERS and EMISAR for snow monitoring in mountainous areas. In: Proc. CEOS SAR Workshop (Toulouse, France, 26-29 October, 1999). Haefner, H. & Piesbergen, J. (1998) Monitoring high mountain snow cover using data fusion techniques. Int. Archives Photogram. Remote Sens. XXXII/7 (Budapest), 350-356. Nagler, T., Rott, H. & Glendinning, G. (1998) SAR-based snow cover retrievals for runoff modelling. In: Proc. Second Workshop on Retrieval of Bio- & Geophysical Parameters from SAR Data for Land Applications (ESTEC, Noordwijk, The Netherlands, 21-23 October 1998), 511-517. ESTEC. Piesbergen, J., Holecz, F. & Haefner, H. (1998) Multisource snow cover monitoring in Eastern Switzerland. In: Proc. IGARSS'98, Seattle, 871-875. Simpson, J. J. & Gobat, J. I. (1995) Improved cloud detection for daytime AVHRR scenes over land. Remote.Sens. Environ. 55(1), 21-49. Srivastava, S. K., Banik, B. T., Adamovic, M., Gray, R., Hawkins, R. K., Lukowski, T. I., Murnaghan, K. P. & Jefferies, W. C. (1999) RADARSAT-1 image quality update. In: Proc. CEOS SAR Workshop (Toulouse, France, 26-29 October, 1999). Yhann, S. R. & Simpson, J. J. (1995) Application of neural networks to AVHRR cloud segmentation. IEEE Trans. Remote Sens. 33(3). of Geosci.