APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT 1

JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION Vol. 44, No. 6 AMERICAN WATER RESOURCES ASSOCIATION December 2008 APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT 1 Amir E. Azar, Hosni Ghedira, Peter Romanov, Shayesteh Mahani, Marco Tedesco, and Reza Khanbilvardi 2 ABSTRACT: Flood forecast and water resource management requires reliable estimates of snow pack properties [snow depth and snow water equivalent (SWE)]. This study focuses on application of satellite microwave images to estimate the spatial distribution of snow depth and SWE over the Great Lakes area. To estimate SWE, we have proposed the algorithm which uses microwave brightness temperatures (Tb) measured by the Special Sensor Microwave Imager (SSM I) radiometer along with information on the Normalized Difference Vegetation Index (NDVI).The algorithm was developed and tested over 19 test sites characterized by different seasonal average snow depth and land cover type. Three spectral signatures derived from SSM I data, namely T19V- T37V (GTV), T19H-T37H (GTH), and T22V-T85V (SSI), were examined for correlation with the snow depth and SWE. To avoid melting snow conditions, we have used observations taken only during the period from December 1-February 28. It was found that GTH, and GTV exhibit similar correlation with the snow depth SWE and are most should be used over deep snowpack. In the same time, SSI is more sensitive to snow depth variations over a shallow snow pack. To account for the effect of dense forests on the scattering signal of snow we established the slope of the regression line between GTV and the snow depth as a function of NDVI. The accuracy of the new technique was evaluated through its comparison with ground-based measurements and with results of SWE analysis prepared by the National Operational Hydrological Remote Sensing Center (NOHRSC) of the National Weather Service. The proposed algorithm was found to be superior to previously developed global microwave SWE retrieval techniques. (KEY TERMS: snow; snow depth; SWE; remote sensing; microwave.) Azar, Amir E., Hosni Ghedira, Peter Romanov, Shayesteh Mahani, Marco Tedesco, and Reza Khanbilvardi, 2008. Application of Satellite Microwave Images in Estimating Snow Water Equivalent. Journal of the American Water Resources Association (JAWRA) 44(6):1347-1362. DOI: 10.1111 j.1752-1688.2008.00227.x INTRODUCTION Understanding seasonal variation of snowcover and snowpack properties is of critical importance for effective management of water resources. According to the Federal Emergency Management Agency, floods are one of the most common hazards in the United States. A re-analysis of the National Weather Service showed that flood damage has been increasing despite local and federal efforts to mitigate floods. Snowmelt is one of the primary reasons for floods. Accurate information on seasonal variation of snowcover and snowpack properties is critical for flood 1 Paper No. JAWRA-07-0022-P of the Journal of the American Water Resources Association (JAWRA). Received February 5, 2007; accepted January 30, 2008. ª 2008 American Water Resources Association. Discussions are open until June 1, 2009. 2 Respectively (Azar, Ghedira, Mahani, and Khanbilvardi), Post Doctoral Research Associate, Research Associate Professor, Associate Professor, and Professor, NOAA-CREST, City University of New York, 137th St and Convent Avenue, New York, New York; (Romanov) Research Scientist, NOAA-NESDIS, Camp Springs, Maryland, and (Tedesco) NASA-Goddard Space Flight Center (E-Mail Azar: eazar@ce.cony.cuny.edu). JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1347 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI prediction and for the effective management of water resources. Current hydrological models predicting snowmelt runoff rely on snowpack measurements made at ground-based meteorological stations. Quite often the density of station network is not sufficient to adequately reproduce the snow cover distribution. Some areas are not covered with surface observations at all. This fact limits the ability to accurately characterize the river runoff and to predict floods. Satellite observations present an important source of information on snow cover properties which can be effectively used to complement traditional groundbased measurements or even substitute them. The launch of Earth Observatory Satellites (EOS) in the mid-20th Century and their capability to observe the earth on large scales encouraged the meteorologists and hydrologists all around the world to find alternatives for traditional methods of estimating snowpack properties. The history of using satellite data for climatological purposes started in 1966 by the launch of National Oceanic and Atmospheric Administration (NOAA) first polar orbiting satellite capable of obtaining visible images of the earth designed to estimate snowcover from space. Satellites operating in the optical wavelength have monitored snowcover over the Northern Hemisphere for more than 40 years (Grody and Basist, 1996). Optical sensors can detect snowcover only during daytime and under cloud-free conditions. In contrast to the visible spectral bands, satellite observations in the microwave do not require daylight and can be used to detect snowcover through clouds. Beside information on the snow cover distribution, satellite microwave instruments offer potential for monitoring physical properties of the snow pack, particularly its water equivalent (SWE) and the snow depth. Snow emission in the microwave domain is highly sensitive to variation of physical prosperities of the snowpack. At frequencies higher than 15 GHz, snow microwave emission tends to decrease as the snowpack thickness increases (Hallikainen, 1984). The radiance measured by microwave sensors is typically converted to corresponding brightness temperature and is expressed in degrees K. Brightness temperature relates surface emissivity (e) to the physical temperature of the object (Ts) (De Seve et al., 1997). In the last three decades a large number of algorithms and techniques have been developed to estimate snow pack properties from satellite observations in the microwave. Chang et al. (1987) proposed a linear relationship between the snow depth (SD) and the brightness temperature difference at 37 and 18 GHz at horizontal polarization SD = 1.59(T18H-T37H). This relationship was established assuming that the density of snow pack and the snow grain size were correspondingly 0.3 g cm 3 and 0.3 mm. The algorithm was applied to global observations of Scanning Multi-channel Microwave Radiometer (SMMR) and Special Sensor Microwave Imager (SSM I) (Foster et al., 1997). Hallikainen (1984) have also used the difference between brightness temperatures at 18 and 37 GHz at horizontal polarization, but related this difference to SWE. Another algorithm utilizing the difference of brightness temperatures at 37 and 19 GHz at vertical polarization was employed by Walker and Goodison (1995) to estimate the snow water equivalent over Canadian Prairies. A modified version of this algorithm was employed by De Seve et al. (1997) to assess snowpack properties over James Bay area in La Grande River watershed, in Quebec, Canada with SSM I data. Tedesco et al. (2004) proposed an Artificial Neural Network technique for the retrieval of SWE from SSM I. They used a multilayer perceptron with various inputs to estimate SWE. The accuracy of snow depth and SWE retrievals with microwave data is generally low. Retrieval errors vary from 70 to 200% depending on a particular area, land cover type, physical conditions of the snow pack, etc (Kelly et al., 2003). Underestimations of SWE often occur due to limited dynamic range of all linear algorithms. Since 2002 global observations of snow cover are also performed with a new generation satellite instrument, Advanced Microwave Scanning Radiometer EOS (AMSR-E). As a prototype AMSR-E global snow depth estimation algorithm, Kelly et al. (2003) introduced an algorithm that combines an empirical snow grain growth model with a densification model that are used to parameterize a constrained Dense Medium Radiative Transfer model suite of snow depth estimates from brightness temperature differences. When compared with snow depth data from station measurements, their algorithm, had an average error of 21 cm, equivalent to 94%. Vegetation is another factor that tends to increase the error of snow depth or SWE retrieval. In order to account for the vegetation effect, Derksen et al. (2004) developed a technique incorporating different linear algorithms for open environments, deciduous, coniferous, and sparse forest cover. The SWE was then calculated as a weighted average of all four estimates, SWE ¼ F D SWE D þ F C SWE Cþ F S SWE S þ F O SWE O ; where (F) is the fraction of each land cover type within a pixel, D, C, S, and O correspondingly represent deciduous forest, coniferous forest, S sparse forest, and O open prairie environments. The effect of forest cover on the emission of snow covered terrain and, thus on the retrieval algorithm JAWRA 1348 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT performance, depends on its type: deciduous forests in winter do not have leaves and thus attenuate microwave radiation to a much lesser extent than coniferous forests. [Therefore, algorithms simply incorporating forest fraction without any account for the forest type (i.e., Foster et al., are not quite correct).] We propose to use Normalized Difference Vegetation Index (NDVI) as an indicator of the forest cover type. In contrast to the vegetation season, NDVI variation in winter is small. NDVI values are largest over evergreen needle-leaf forests, and decreases over mixed forests and deciduous broad-leaf forests. Over snow-covered non-forested areas NDVI even reaches negative values because of higher reflection of snow in the visible spectral band than in the near-infrared. In this study, we developed and tested a new algorithm for estimating snow depth and SWE from satellite observations in the microwave. The primary focus was on the Great Lakes area. The technique incorporates a two-stage algorithm and uses NDVI to account for vegetation effects. The algorithm was tuned using surface observations. The accuracy of the algorithm was evaluated through the comparison of satellite retrievals with surface observations and with the output of a physicalbased snowpack model developed and run operationally at National Operational Hydrologic Remote Sensing Center (NOHRSC). In the first section of this paper, we present the study area location and land cover characteristics, satellite data, as well as snow observations and modeled data. The second section describes the methodology to evaluate capability of microwave data in retrieving snowpack properties in the study area. The third section discusses the results that were used to develop a new model. The last section describes the development of the new algorithm and evaluates its performance over the Great Lakes area. the behavior of satellite measured microwave radiations with respect to snow over various land cover types. SSM I Data The SSM I passive microwave radiometer has seven channels operating at five frequencies (19, 35, 22, 37.0, and 85.5 GHz) and dual- polarization (except at 22 GHz which is vertical polarization only) (Table 1). The sensor spatial resolution varies for different channels frequencies. In this study, the Scalable Equal Area Earth Grid EASE-Grid SSM I products distributed by National Snow and Ice Data Center (NSIDC) were used (Brodzik and Knowles, 2002). In SSM I EASE-Grid, all channels below 85 GHz are re-sampled to footprint size of 19 GHz beam but the sample spacing is slightly more than 25 km (25.06 km) for all the channels (NSIDC) (Armstrong et al., 1994). The EASE-Grid SSM I data are available in global cylindrical, and azimutal equal area. Since our study area is between 45 N and 49 N, we have used the Northern Hemisphere Azimuthal Equal-Area EASE-Grid. EASE-Grids aspect ratio is about 1.17:1 at 45 N as compared with cylindrical aspect ratio which is 1.50 for 45 N, making Azimuthal projections more desirable (Brodzik and Knowles, 2002). The study area, located between 41 and 49 N and 87 and 98 W, is covered by 28 X 35 (980) EASE-Grid pixels. Normalized Difference Vegetation Index Proposed by Rouse et al. (1973), NDVI is widely used to characterize vegetation cover. NDVI is defined as a difference between reflectance in visible red and near-infrared spectral bands divided by their sum STUDY AREA AND DATA USED ðndvi ¼ðNIR VISÞ=ðNIR þ VISÞÞ: The selected study area is located west of Great Lakes between 41 N and 49 N and 87 W and 98 W covering parts of Minnesota, Wisconsin, and Michigan. The area covers hundreds of water sheds in three major basins of Great Lakes basin, Souris-Red River basin, and upper Mississippi River basin. The study area has different land cover types ranging from bare land and grass land to deciduous and needle-leaf forests. Diversity of land cover type was among the major reasons for selection of Great Lakes area for this study in order to analyze TABLE 1. SSM I Channels, Polarizations, and Resolutions. Frequency (GHz) Polarization Footprint Along Track (km) Footprint Across Track (km) 19.35 Vertical 69 43 19.35 Horizontal 69 43 22.235 Vertical 50 40 37 Vertical 37 28 37 Horizontal 37 29 85.5 Vertical 15 13 85.5 Horizontal 15 13 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1349 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI Live green plants appear relatively dark in the visible and relatively bright in the near-infrared and thus exhibit high NDVI values (Gates, 1980). Soil and bare land have lower NDVI which even becomes negative if the land is covered by snow. The NDVI data for this study were obtained from the NOAA NASA Pathfinder Advanced Very High Resolution Radiometer dataset which is distributed by Goddard Space Flight Center (DAAC). The NDVI data are extracted from a global 10-day composite image for January 21-31 in 1994. The composite images are derived by from images in a 10-day period with minimum cloud coverage. To facilitate the comparison and matching of the two datasets (NDVI and SSM I) NDVI data were re-sampled and projected to the EASE-Grid projection at 25 km spatial resolution. Normalized Difference Vegetation Index has a seasonal pattern meaning that it increases during spring and summer and decreases during winter. The winter NDVI tend to be much lower than summer for all types of land cover. Also, NDVI variation during the winter season is very limited. Maximum winter NDVI is generally observed over evergreen needle-leaf forests, which decreases over mixed forests and deciduous broad-leaf forests. Over grass land and bare land which is covered by snow NDVI becomes negative. On the other hand, microwave scattering is related to land cover. One of the sources of error in estimating SWE from microwave data is attenuation of microwave scattering over the forested areas (evergreen and mixed forests). By using winter NDVI data, the attenuation effect can be estimated well. Ground-Based Snow Measurements Surface observations performed at first-order and US Cooperative Network Stations were obtained from National Climatic Data Center (NCDC). There are 681 stations within the study area but they are not uniformly distributed. Most stations are located in the vicinity of densely populated areas close to the lake. To develop the algorithm and to evaluate its performance we have used measurements made at 19 specifically selected test sites. Each test site is size of an EASE-Grid pixel (25 km 25 km). For the test sites with more than one station, the observations were averaged. Table 2 lists geographical location of the selected test sites and their NDVI characteristics including the mean and standard deviation within corresponding EASE-Grid, 25 km resolution cells. The mean and standard deviation are derived based on difference between spatial resolution of EASE-Grid (25 km 25 km) and NDVI (8 km 8 km). Model Data Snow products generated by the Snow Data Assimilation System (SNODAS) of NOAA National Weather Service s National Operational Hydrologic Remote Sensing Center (NOHRSC) are available beginning October 2003. SNODAS presents a physically based, spatially distributed energy and mass balance model which incorporates ground-based observations of snow depth, air-borne measured gamma radiations, and downscaled output from regional Numerical Weather Prediction as input (NOHRSC, 2004).The output of the system includes fields of snow depth, snow water equivalent, snow melt, and a number of snowpack characteristics generated at 1 km spatial and hourly temporal resolution. In order to match the SSM I and SWE datasets, we converted the resolution of SNODAS- SWE data to 25 km. Although, NORHSC snow data are produced from incorporating data from intense network of snow reporting stations and have very high resolution, but there are some limitations associated such as dependency on air-borne gamma measurements which are limited and costly. In addition, these data are available only over United States and their accuracy is not well evaluated. In order to evaluate the consistency of the SNO- DAS-SWE data with NCDC snow depth observations, we calculated the correlation coefficient between Test Site TABLE 2. Coordinates of Selected Pixels Along With NDVI Values, Each EASE-Grid Pixel Contains 3 3 NDVI Pixels. SSM I EASE-Grid Pixels NDVI Latitude Longitude Center Mean Standard Deviation 1 42.33 )93.62 )0.040 )0.033 0.008 2 42.89 )91.97 )0.040 )0.038 0.005 3 43.63 )91.43 )0.032 )0.025 0.012 4 44.14 )90.57 0.136 0.121 0.020 5 44.39 )89.12 0.000 0.032 0.016 6 45.12 )89.11 0.016 0.034 0.018 7 46.07 )88.19 0.272 0.219 0.029 8 45.59 )88.21 0.192 0.237 0.024 9 46.09 )88.79 0.256 0.268 0.024 10 46.80 )88.46 0.304 0.196 0.026 11 46.80 )88.16 0.176 0.234 0.023 12 46.83 )89.69 0.248 0.247 0.026 13 45.36 )91.18 0.032 0.038 0.022 14 45.56 )92.68 )0.024 0.023 0.015 15 47.26 )92.78 0.168 0.123 0.018 16 48.01 )91.88 0.160 0.212 0.025 17 47.92 )94.08 0.056 0.079 0.020 18 48.40 )95.99 0.056 0.035 0.017 19 47.47 )97.47 )0.024 )0.026 0.007 JAWRA 1350 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT NCDC snow depth observations vs. SNODAS-SWE products over the 19 selected test sites. The results, shown in Table 1, indicate satisfactory correlations between NCDC snow depth observations and SNO- DAS-SWE. It is observed that in some test sites such as test Sites 4, 5, 11, and 12 the correlation coefficient has decreased significantly as compared with other test sites. This decrease can be associated to average seasonal snow fall, and existence of water bodies within the test site (Azar, 2006). METHODOLOGY In this study, we are proposing a new algorithm for SWE snow depth estimations which is tuned for the Great Lakes area and its land cover characteristics (Figure 1). In order to develop the algorithm, first we examine microwave observations and their potentials for estimating snow pack properties in the Great Lakes area. Then, we investigate the effect of incorporating NDVI data in microwave-based snow estimates. Finally, using the results of the analysis, we propose a new algorithm that incorporates both SSM I and NDVI data and evaluate the performance of the new algorithm over the Great Lakes area. As it was mentioned earlier the study area includes various types of land cover (Figure 1). The area is covered by 28 by 35 SSM I EASE-Grid pixels. Nineteen test sites were selected; each of the sites had a size of an EASE-Grid pixel (Table 1). In selecting particular test sites we considered three criteria: 1 Availablility of snow depth measurements. 2 Covering different types of land cover throughout the study area. 3 Sufficient distance between the test site and the border of the lake where SSM/I measurements could be affected by water. Considering the mentioned criteria, there was only limited number of locations available to be selected as test sites (Figure 1). To avoid wet snow conditions only the data from December 1 to February 28 were considered. Wet snow has a negligible scattering signal and needs to be excluded from the retrieval (Goodison- Walker 1993). Three datasets (containing 90-day information) were derived for each winter seasons 2001-2004. Evaluation of the Microwave SSM I Channels In order to evaluate the potentials of SSM I data in snow depth SWE estimations in Great Lakes area, we investigated the behavior of three SSM I scattering signatures with respect to snow depth and water equivalent. The first scattering signature, GTH (19H-37H) is the gradient brightness temperatures (Tb) between SSM I channels in 19 and 37 GHz in horizontal polarization. This signature was used by Chang in his global snow depth retrieval algorithm (Chang et al., 1987). The second scattering signature was used by Goodison-Walker to estimate SWE in Canadian prairies. Similar to Chang s algorithm, the signature is defined as gradient of brightness temperatures (Tb) between 19 and 37 GHz but in vertical polarization, GTV (19V-37V) (Goodison-Walker 1995). The third scattering signature was defined for shallow snow identification and estimations as the difference between channels 22 and 85 GHz in vertical polarization, SSI (22V-85V). The fact that 85 GHz is the most sensitive channel to snow can make SSI FIGURE 1. Land Cover Image Extracted From USGS National Atlas of Land Cover Characteristics (USGS, seamless data, last modified July 2003). JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1351 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI an excellent signature for snow identification but for SWE and snow depth estimations this channel (85 GHz) is bounded to saturation problem. The box-whiskers plot of GTH and GTV values during the winter seasons are illustrated (Figure 2). The negative outliers in the box plots are due to either sensor or data processing errors which need to be eliminated. The radiation in 19 GHz must be higher than the scattering radiation in 37 GHz because of higher absorption of emitted radiation in 37 GHz over a snow covered surface. In addition, GTV mean ranges from 5 to 15 for all the pixels except for the test Site 12 which is very close to the lake. In test Site 12, the mean of GTV is around )5. The relatively large negative value for GTV is due to contamination of the scattering signals from land by that of water. The SSM I sensors have different spatial resolutions for different channels. In EASE- Grid data, the signal is averaged to the footprint of l9 GHz (69 km 43 km), using Backus-Gilbert technique and then re-sampled to 25 km 25 km (Armstrong et al., 1994). Thus, although, the grid (test Site 12) is not located in the water but its scattering values and consequently its brightness temperatures are contaminated by scattering signals of water. Evaluation of the Microwave SSM I Channels for Snow Estimations FIGURE 2. Box-Whiskers Plot of GTV, GTH, and SSI for Winter 2003-2004. After eliminating the negative outliers, a 3 year time series of GTV and snow depth for each of the test sites was produced. Figure 3 illustrates SSM I signature of GTV and snow depth at Site 9. The plot shows that as snow depth increases during the winter the GTV increases. This is because of the high sensitivity of channel 37 GHz to snow. Contradictory to the northern test sites, those test sites located in the more southerly area, do not show a consistent seasonal pattern for snow depth and GTV (Figure 4). Figures 3 and 4 illustrate high correlations between snow depth and GTV for northern test sites and low correlation for southern test sites. In order to quantitatively evaluate the capability of SSM I data in estimation of snow depth, the correlation coefficients between the three SSM I scattering signatures, GTV (19V-37V), GTH (19H-37H), SSI (22V-85V) vs. snow depth was derived (Table 3). To visualize the results presented in Table 3, the variation of the correlation coefficients is presented in Figure 5. SSM I signatures at Sites 11, 12, and 13 does not show correlations with snow depth since the land scattering is affected by the lake scattering since those pixels are close to the lake. In fact, JAWRA 1352 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT FIGURE 3. Three Year Time Series of GTV (19-37V) vs. Snow Depth for Point 9. FIGURE 4. Three Year Time Series of GTV (19-37V) vs. Snow Depth for Point 2. TABLE 3. Correlations of Snow Depth vs. SSM I Signatures GTH (19H-37H), GTV (19V-37V), and SSI (22V-85V). Test Sites Winter 01-02 Winter 03-02 Winter 03-04 GTH GTV SSI GTH GTV SSI GTH GTV SSI 1 0.00 0.00 0.40 0.20 0.10 0.30 0.52 0.60 0.70 2 0.10 0.30 0.80 0.10 0.30 0.80 0.45 0.40 0.63 3 0.13 0.20 0.50 0.05 0.10 0.50 0.20 0.15 0.20 4 0.20 0.20 0.20 0.05 0.12 0.25 0.60 0.55 0.70 5 0.00 0.00 0.10 0.00 0.10 0.11 0.10 0.10 0.33 6 0.50 0.55 0.10 0.00 0.11 0.12 0.22 0.20 0.10 7 0.70 0.68 0.00 0.00 0.00 0.00 0.25 0.30 0.05 8 0.55 0.53 0.00 0.00 0.00 0.00 0.00 0.00 0.00 9 0.71 0.70 0.20 0.20 0.30 0.20 0.15 0.30 0.20 10 0.30 0.50 0.00 0.00 0.10 0.10 0.05 0.10 0.00 11 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 12 0.00 0.00 0.20 0.20 0.35 0.21 0.15 0.35 0.20 13 0.35 0.35 0.00 0.00 0.00 0.00 0.00 0.15 0.00 14 0.40 0.30 0.40 0.33 0.30 0.40 0.35 0.25 0.45 15 0.50 0.50 0.10 0.35 0.45 0.10 0.35 0.50 0.1 16 0.60 0.60 0.30 0.40 0.50 0.30 0.45 0.53 0.30 17 0.81 0.80 0.10 0.80 0.80 0.00 0.80 0.80 0.05 18 0.92 0.90 0.30 0.50 0.50 0.35 0.50 0.50 0.40 19 NA NA NA NA NA NA NA NA NA FIGURE 5. Correlations of Snow Depth vs. SSM I Signatures GTV (19V-37V), GTH (19H-37H), and SSI (22V-85V) for Various Test Sites (TS) for Winter Seasons 01-02, 02-03, 03-04. JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1353 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI FIGURE 6. Variation of Correlations of SWE vs SSM I Scattering Signatures for Various Points for Winter 03-04. those pixels are affected by the lake scattering in 19 GHz channel (69 km) but are not affected in 37 GHz (50 km) and 85 GHz (15 km). It is also observed that SSI shows higher correlations with snow depth in test Sites 1-4 where the NDVI values are below zero. On the other hand, GTV and GTH show higher correlations with snow depth where the NDVI values are high. Similar approach was taken for analyzing the behavior of SWE products from NOHRSC with respect to SSM I scattering signatures (Figure 6). The correlation coefficients between SWE and SSM I signatures follow the same pattern except those are slightly higher that correlations with snow depth. The scatter plots of SWE vs. the three SSM I signatures (GTV, GTH, SSI) have been produced for the all the test sites. Figure 7 illustrates the variation of SWE vs. scattering signatures for selected test sites (2, 9, and 18). The scatter plot of SSM I signatures vs. SWE indicate that the slope of regression line is different for various test sites considering their land cover type which is represented by NDVI values. These variations in the regression slopes originate from scattering attenuation over the forested areas. Variation of regression FIGURE 7. Scatter Plots of SSM I Scattering Signatures Signature [(GTH (19H-37H), GTV (19V-37V), and SSI (22H-85H)] vs. SWE (SNODAS) for Winter 2003-2004. JAWRA 1354 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT slope over is a source of error for those microwavebased SWE estimating algorithms which do not take vegetation in to account. Then, in order to reduce the effect of forest scattering attenuation we investigated the possibility of using NDVI value for estimating the regression slope and ultimately SWE and snow depth. Table 4 shows how the correlation coefficient, the slope and intercept of the regression line from scattering signatures of SSM I channels varies with respect to maximum SWE received by the test sites. Similar to snow depth behavior with respect to SSM I, in test sites that receive <60 mm (1-4) of SWE, SSI shows higher correlation with SWE. On the other hand, for test sites which receive more than 80 mm (7-9, 15-17, 19) of SWE, GTV has the dominant correlation with SWE. In test sites where maximum SWE is between 60 and 80 mm, the behavior of both SSI and GTV is similar, non shows very high correlation with SWE. On these test sites, either of the signature can be used but to find the optimum answer, other criteria need to be used which will be described in the algorithm development section. variation, represented by NDVI, and microwave scattering, represented by regression slopes for different test sites. Azar et al. (2006) conducted a research on NDVI variation with respect to microwave scattering in Great Lakes area revealing that the microwave scattering attenuation by vegetation over a snowpack is highly correlated with the NDVI computed over the same area. Figure 8 illustrates the variation of the slope regression slopes between GTV and SWE, and NDVI for all the sites during winter 2003-2004. It is observed that the regression slope is higher for test sites with high NDVI and it decreases for the test sites with low NDVI indicating a high correlation between the two parameters. The scatter plots of the regression slopes (derived from GTV) vs. NDVI are illustrated in Figure 9. The plot is drawn only for those points (test sites) which the NDVI is higher than zero. It was shown that for the NDVI less than zero the SSI has higher correlation with snow depth (Azar et al., 2006). Evaluation of NDVI Variations Over Different Test Sites With Microwave SSM I Channels for Snow Estimations Scatter plots of SWE vs. SSM I scatterings showed that there might be a connection between land cover TABLE 4. Variation of Correlations and Regression Slopes With Maximum Seasonal SWE for the Selected Test Sites. Test Site Max SWE (mm) Correlation Coefficient SSI Slope SSI Intercept SSI Correlation Coefficient GTV Slope GTV 1 60 0.7 1.9 0.7 0.3 2.2 2 35 0.84 )3 0.84 0.29 1.9 3 40 0.77 )2.4 0.77 0.35 2.1 4 55 0.7 )2.5 0.70 0.31 2.2 5 70 0.67 )4.4 0.67 0.64 4.4 6 80 0.47 NA NA 0.67 6 7 130 0.50 NA NA 0.8 9.4 8 120 0.16 NA NA 0.55 7.9 9 175 0.47 NA NA 0.77 15 10 NA NA NA NA 0.62 5.5 11 NA NA NA NA NA NA 12 NA NA NA NA NA NA 13 75 0.52 NA NA 0.5 5.3 14 50 0.67 )7.7 0.67 0.44 2.4 15 95 0.28 NA NA 0.6 3.8 16 170 0.38 NA NA 0.68 9.3 17 90 0.1 NA NA 0.84 4.7 18 50 0.65 )4.8 0.6 0.83 1.8 19 85 0.6 NA NA 0.71 3.9 FIGURE 8. Variations of the Slope of the Regression Lines in the Scatter Plots With NDVI for the Test Sites for Winter 03-04, SWE vs. GTV (up), SD vs. GTV (down). JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1355 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI FIGURE 9. Scatter Plots of the Slope of the Regression Lines vs. NDVI. RESULTS AND DISCUSSION For all the sites GTV and GTH show similar behavior with respect to snow depth. In other words, the difference between vertically and horizontally polarized signatures is negligible in terms of correlations with snow depth. Contrary to GTV and GTH, SSI has a different pattern. It has the dominant correlation for test sites 1-4 but for sites located in high latitudes GTV becomes the dominant. This is because of the saturation of the 85 GHz channel used in SSI over a deep snow pack. However, SSI can be used to identify and to estimate SWE over shallow snow. In case of SWE and SSM I signatures, Figure 6 illustrates the correlations between SWE and different SSM I spectral signatures. For test Sites 1-4 SSI has the higher correlation but for the sites (9-7, 15-17, and 19) GTV and GTH show better correlations with SWE. The rest of the test sites are acting in between. Figure 6 also shows that the correlations between SWE and scattering signatures are higher than those for snow depth. The higher correlation between SSM I and SWE obtained from NOHRSC can be explained by that fact that NOHRSC SWE data are produced from combining station observation with output of weather prediction models that have low spatial resolution similar to SSM I data. The correlation coefficients between snow depth and scattering signatures follow a consistent pattern for all the winter seasons. The only inconsistency is for sites 6-10 for winter 01-02 which can be explained by the snow received by those test sites during the winter season. According to NCDC snow depth data, test Sites 6-10 received more snow in winter 01-02 compare with winters 02-03 and winter 03-04 when those test sites received less snow. No correlations between scattering signatures and snow depth and water equivalent were observed at close to Great Lakes (11, 12, 13). This is due to the contamination of land scattering signal by scattering signal from water bodies resulting from the re-sampling of SSM I to EASE-Grid by using Backus-Gilbert interpolation technique. Figure 7 shows the scatter plots of SSM I signatures vs. SWE (SNODAS) and SSM I signatures vs. snow depth (stations) for winter 2003-2004 in different test sites. Regression lines for each graph have various slopes and intercepts (Table 3). The slopes increase over highly vegetated areas. This demonstrates that having a single linear algorithm (e.g., Chang or Goodison-Walker) may not be appropriate for snow depth or SWE in a variety of environmental and geographical conditions. Figures 8 show that the regression slopes are highly correlated with NDVI values. Over test sites with high NDVI, highly vegetated, the regression slope tends to increase. Then, having the NDVI value for a particular area, we can derive the regression slope (Figure 9). In short, GTH and GTV are similar and both show high correlations with snowpack properties. Both of them are not reliable indicators in vicinity of open water (69 km). In addition, SSI is more sensitive to snow variation in particular over shallow snow. Finally, NDVI can be used to derive the regression slope over different areas. NEW ALGORITHM DEVELOPMENT AND EVALUATION In this section, propose a new algorithm based on the findings in the previous sections. The proposed algorithm will be evaluated over the whole study JAWRA 1356 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT area. It is also evaluated temporally over the selected test sites. The results are compared with that of other algorithms such as Chang s global algorithm and Goodison-Walker regional algorithm. Development of the New Algorithm Considering the analysis and results detailed in the previous sections, we propose: 1. Using SSI (22V-85V) for shallow snow estimations. 2. Using GTV (19V-37V) for deeper SWE snow depth estimations along with the corresponding regression slope. 3. Using NDVI value to derive the regression slope for GTV The decision-tree algorithm is illustrated in Figure 10. Where SWE is the snow water equivalent in mm, GTV, and SSI are SSM I spectral scattering signatures. Winter time NDVI was obtained from a FIGURE 10. Algorithm to Estimate Snow Water Equivalent (SWE) Using SSM I Data in the Great Lakes Area. 10-day composite image for January 1994. A and B are derived from the scatter plots of regression slope and NDVI. Coefficients C and D are determined from the scatter plots of SWE vs. SSI using the average of the best fitted line to the scatter plots. The Values of coefficients A, B, C, and D entering the above formula were found equal to 37, 1.8, 1.03, and )3.27 respectively. Algorithm Validation The new algorithm was examined over by the whole dataset of matched satellite retrieval and SWE estimates in Great Lakes region. All of the test site, the new algorithm showed significant improvement in reducing the RMSE as compared with Goodison- Walker or Chang s algorithm. Figure 11 shows the results obtained with the new algorithm over test Site 10 as compared to Goodison-walker and Chang algorithms. The tests Site 10 is located in latitude 46.8N and longitude )88.46W in the area covered with mixed forest. The results indicate decrease of Root Mean Square Error (RMSE) over the test Site 10 which is the result of introducing NDVI value into the equations. In order to derive snow depth, the estimated SWE values were multiplied by the average snow density (0.23 gr cm 3 ). The new algorithm was validated for the three winter periods (December 01-February 28) 2001-2004. In winter 2003-2004, both SWE (from NOHRSC) and snow depth (from NCDC) data were available. For winter seasons 2001-2002 and 2002-2003, only snow depth data was available. The results indicate 28 mm, 33 mm of reduction in RMSE for SWE estimations as compared with results from Chang and Goodison-Walker algorithms consecutively. Similarly, estimated snow depth with the new algorithm is more accurate that the other two algorithms. The source of error for Chang and Goodison-walker algorithms is mainly underestimation of SWE and snow depth (Figure 11). This underestimation is due to attenuation of microwave scattering in the forested area which is reflected in reduction of brightness temperature (Tb). By introducing NDVI to the algorithm equation, the new algorithm takes vegetation effect into account for snow estimations. Besides the temporal validation, the new algorithm was spatially validated for the whole study area (Latitudes: 41-49N and Longitudes: )87W to )98W) excluding the areas covered by the lakes. The EASE-Grid pixels covered or in the vicinity of the lakes, were filtered out. There were 11 days (3 days December, 4 days January, and 4 days February) with full coverage of SSM I data in winter JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1357 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI FIGURE 11. Scatter Plots of Estimated vs. True Snow Depth SWE for Different Algorithms for Test Site 10 (Lat = 48.6N, Lon = )88.46W, and NDVI = 0.2). JAWRA 1358 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT FIGURE 12. Comparison of Spatial Distribution of Estimated SWE by Various Algorithms With Ground Truth Data for January 25, 2004 for the Study Area (Lat: 41N-49N and Lon: )87W to )98W). FIGURE 13. NDVI Image and Results of Estimated SWE vs. Ground Truth for January 25, 2004. JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1359 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI FIGURE 14. Scatter Plots of Estimated SWE by Chang and Goodison-Walker Algorithms vs. Ground Truth for January 25, 2004. 2003-2004. The ground truth data was obtained by averaging NOHRSC SNODAS dataset. The new algorithm was used to estimate SWE spatial distribution over the study area. Figure 12 shows the ground truth and estimated SWE for January 25, 2004. The NDVI image of the study area (Figure 13) shows higher values of NDVI around the lake. This is the area that both Chang and Goodison- Walker algorithms highly underestimate the SWE (Figure 12). In contrast, the algorithm can estimate SWE in the area in the vicinity of the lake with much higher accuracy (Figures 12 and 13). The calculated RMSE and correlation coefficient (R 2 ) are shown for all the three algorithms. The use of NDVI in the new algorithm results in a decrease of the RMSE and the increase of the correlation coefficient. It also increases the range for the estimated SWE. Table 4 demonstrates a consistent improvement in the accuracy of the estimated SWE for the winter season of 2003-2004. In average, compared with Chang global algorithm, the correlation coefficient is improved about 0.20 and the RMSE is decreased about 4 mm of SWE over the study area. For all days, application of the new developed algorithm results in the highest correlation coefficient between SSM I and SWE. At the same time, the RMSE of SWE derived with the new algorithm is lower for all days but those in February (Table 5). There is a decreasing trend of in correlations and increasing trend in SWE in February. The most probable reason for this trend is snow melt. In TABLE 5. Variations of RMSE and Correlation Coefficients for Selected Days in Winter 2003-2004. Days Algorithm Chang Correlation Coefficients GW RMSE (mm) NEW Alg. Chang GW NEW Alg. December 6, 2003 0 0 0.29 22 13 11 December 13, 2003 0.35 0 0.30 21 14 12 December 20, 2003 0 0 0.30 33 20 15 January 4, 2004 0.47 0.39 0.69 22 19 17 January 11, 2004 0.10 0.12 0.59 21 16 14 January 18, 2004 0.33 0.28 0.67 28 25 18 January 25, 2004 0.52 0.50 0.70 25 25 19 February 1, 2004 0.5 0.46 0.60 30 31 32 February 8, 2004 0.32 0.20 0.48 37 38 47 February 16, 2004 0.30 0.23 0.43 42 52 54 February 23, 2004 0.30 0.12 0.48 55 54 54 February, the study area and especially its southern part experienced several melt and refreeze of snow. The higher brightness temperature and reported surface temperatures over the study area supports the existence of wet snow for those days (Tedesco et al., 2006). Estimates of snow depth and SWE with satellite observations in microwave become practically impossible when snow is wet. The results presented in Figure 15, confirms the existence of the wet melting snow for the days in February. There are many points on the vertical axis of the scatter plots of microwave estimated SWE vs. SWE from NOHRSC which indicate the snow is not detected by the microwave-based algorithm. JAWRA 1360 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION

APPLICATION OF SATELLITE MICROWAVE IMAGES IN ESTIMATING SNOW WATER EQUIVALENT FIGURE 15. Validation of the New Algorithm Over the Great Lakes Area on Different Days (dates are selected based on full coverage of the study area by SSM I data). CONCLUSIONS A new algorithm was developed to estimate SWE using SSM I scattering Signatures and NDVI over Great Lakes area of United States. Current linear algorithms such as Goodison-Walker and Chang algorithms are not sufficient for accurate estimations of SWE in Great Lakes area. In order to resolve this problem three winter seasons were studied. SSM I data with corresponding snow depth, and snow water equivalent (SWE) were used to examine the sensors response to the changes in snow pack properties. SSM I response in GTV (19V-37V), GTH (19H-37H), and SSI (22V-85V) to snow depth or water equivalent changes were analyzed. In order to minimize wet melting snow conditions, in which microwave signatures cannot be used to estimate SWE snow JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION 1361 JAWRA

AZAR, GHEDIRA, ROMANOV, MAHANI, TEDESCO, AND KHANBILVARDI depth, only the periods between December 1 and February 28 were considered. In more southerly areas where snow is mostly shallow, SSI has the highest correlation with SWE. In northern part of the study area, GTV and GTH are better estimators of SWE. Also, the scatter plots of SWE vs. GTV and GTH shows that the slope of the regression line between the spectral signatures and SWE varies with location. This variation of the slope was found to be correlated to NDVI and was employed in the new algorithm for estimating SWE using SSM I data over the Great Lakes area. The new algorithm was spatially validated for the whole study area, excluding the areas covered by the lakes. The use of NDVI in the new algorithm results in a decrease of the RMSE and the increase of the correlation coefficient. It also increases the range for the estimated SWE. In average, compared with Chang global algorithm, the correlation coefficient is improved about 0.20 and the RMSE is decreased about 4 mm of SWE over the study area. ACKNOWLEDGMENTS The authors express their gratitude to Meteorological Service of Canada (MSC). Thanks to NOHRSC and NSIDC for providing SNODAS-SWE and SSM I dataset. This study was supported and monitored by National Oceanic and Atmospheric Administration (NOAA) under Grant NA06OAR4810162. The views, opinions, and findings contained in this report are those of the author(s) and should not be construed as an official National Oceanic and Atmospheric Administration or U.S. Government position, policy, or decision. LITERATURE CITED Armstrong, R.L., K.W. Knowles, M.J. Brodzik, and M.A. Hardman, 1994, updated current year. DMSP SSM I Pathfinder Daily EASE-Grid Brightness Temperatures, [List Dates of Data Used. Boulder, Colorado USA: National Snow and Ice Data Center. Digital media. Azar, A.E., 2006. Application of Satellite-Borne Microwaves in Estimation of Snowpack Properties, PhD, City University of New York, Graduate Center, New York, New YorK. pp. 137-153. Azar, A.E., S. Mahani, H. Ghedira, P. Romanov, and R. Khanbilvardi, 2006. Time Series Analysis and Algorithm Development for Estimating SWE in Great Lakes Area Using Passive. 63rd Annual Meeting ESC, Newark, Delaware, June 7-9, 2006. Brodzik, M.J. and K.W. Knowles, 2002. EASE-Grid: A Versatile set of Equal-Area Projections and Grids. In: Discrete Global Grids, M. Goodchild, editor. National Center for Geographic Information & Analysis, Santa Barbara, California. Chang, A.T.C., J.L. Foster, and D.K. Hall, 1987. Nimbus-7 SMMR Derived Global Snowcover Parameters. Annals of Glaciology 9:39-44. De Seve, D., M. Bernier, J.P. Fortin, and A. Walker, 1997. Preliminary Analysis of Snow Microwave Radiometry Using the SSM I Passive Microwave Data: The Case of La Grande River Watershed (Quebec). Annals of Geology 25. Derksen, C., R. Brown, and A. Walker, 2004. Merging Conventional (1915-92) and Passive Microwave (1978-2002) Estimates of Snow Extent and Water Equivalent Over Central North America. Journal of Hydrometeorology 5:850-861. Foster, J.L., A.T. Chang, and D.K. Hall, 1997. Comparison of Snow Mass Estimation From a Prototype Passive Microwave Snow Algorithm, a Revised Algorithm and Snow Depth Climotology. Remote Sensing of Envirionment, 62:132-142. Gates, D.M., 1980. Biophysical Ecology, Springer-Verlag, New York, New York, 611 p. Grody, N. and A.N. Basist, 1996. Global Identification of Snowcover Using SSM I Measurements IEEE Transaction on Geosciences and Remote Sensing, 34(1):237-249. Hallikainen, M.T., 1984. Retrieval of Snow Water Equivalent From Nimbus-7 SSMR Data: Effect of Land Cover Categories and Weather Conditions. IEEE Journal of Oceanic Engineering 9(5):372-376. Kelly, R.E., A.T. Chang, Tsang. Leung, and J.L. Foster, 2003. A Prototype AMSR-E Global Snow Area and Snow Depth Algorithm. IEEE Transactions on Geoscience and Remote Sensing 41(2):230-242. NOHRSC, (National Operational Hydrologic Remote Sensing Center), 2004. Snow Data Assimilation System (SNODAS) Data Products at NSIDC. National Snow and Ice Data Center. Digital Media, Boulder, Colorado. Rouse, J.W., R.H. Haas, J.A. Schell, and D.W. Deering, 1973. Monitoring Vegetation Systems in the Great Plains With ERTS. Third ERTS Symposium, NASA SP-351 I, pp. 309-317. Tedesco, M., E.J. Kim, A.W. England, R.D. De Roo, and J.P. Hardy, 2006. Brightness Temperatures of Snow Melting Refreezing Cycles: Observations and Modeling Using a Multilayer Dense Medium Theory-Based Mode. IEEE Transaction on Geosciences and Remote sensing 44(12):3563-3573. Tedesco, M., P.J. Takala, M. Hallikainen, and P. Pampaloni, 2004. Artificial Neural Network-Based Techniques for the Retrieval of SWE and Snow Depth From SSM I Data. Remote sensing of Environment 90. Walker, A.E. and B. Goodison, 1995. Canadian Development and Use of Snow Cover Information From Passive Microwave Satellite Data. In: Passive Microwave Remote Sensing of Land- Atmosphere Interactions, B.J. Choudhury, Y.H. Kerr, E.G. Njoku, and P. Pampaloni, editors. The Netherlands, VSP BV 245-262. http://daac.gsfc.nasa.gov/interdisc/readmes/pal_ndvi.shtml#500, accessed December 2006. http://seamless.usgs.gov/website/seamless/viewer.php, accessed December 2006. JAWRA 1362 JOURNAL OF THE AMERICAN WATER RESOURCES ASSOCIATION