SIR-C/X-SAR Investigations of Snow Properties in Alpine Region

Seasonal snow cover is an important component in investigations of land-surface climate and hydrology. It is critical to the radiation and water balances. Snow and ice play important interactive roles in regional climates, because snow has a higher albedo than any other natural surface. Snow causes a seasonal variation in surface albedo, and it has a low thermal conductivity, so it insulates the soil surface from large, rapid temperature fluctuations. Of hydrological significance, snow stores water during the winter season and releases it in spring and summer. Over major portions of the middle and high latitudes, and at high elevations in the tropical latitudes, snow and alpine glaciers are the largest contributors to runoff in rivers and to ground-water recharge. Snow hydrologists are confronted with the dual problems of estimating both the quantity of water held by seasonal snow packs and time of snow melt.

In hydrological research and operations, modeling and forecasting of snow-melt runoff requires timely information about snow properties and their temporal and spatial variability. The most common use of remote sensing in snow studies is to monitor snow-covered area. These efforts have been carried one step further by including satellite, from visible and near-infrared sensors, derived measurements of snow-covered area as an index in a snow-melt runoff model. In the visible, near-infrared, and shortwave-infrared spectrum, however, the sensors cannot see through clouds, which are pervasive in some regions and may prohibit the timely acquisition of information needed for both operational and research applications. In addition, the measurements from those sensors are insensitive to snow depth and free liquid water in snow pack because radiation does not penetrate the snow pack and there is no significant difference in dielectric properties between ice and water. On the other hand, active microwave remote sensing has long promised the advantages of (1) sensitivity to many snow parameters that snow hydrologists use, especially snow water equivalence and free liquid water content in the snow pack because of the large dielectric contrast between ice and water in the microwave spectrum; (2) ability to provide day or night imaging in all weather; and (3) a spatial resolution compatible with the topographic variation in alpine regions.

The objectives of this research project is to develop algorithms and to examine the capabilities of SIR­C/X­SAR (Shuttle Imaging Radar-C and X-Band Synthetic Aperture Radar) to measure the snow properties. They are (1) mapping snow covered area; (2) inferring snow wetness; and (3) estimation snow water equivalence that are necessary for the calculation of the snow pack mass balance. This document reviews our accomplishments in the NASA/JPL-supported project: SIR-C\X-SAR investigations of snow properties in alpine regions. In addition to the studies on estimation of snow properties, wehave also developed an algorithm to infer the bare surface soil moisture and surface roughness parameter, co-operating with the NASA/GSFC, which will be summarized in the last session of this report. The References section contains a bibliography of publications which are related to this project.

I. Field Experiments

During two SIR-C/X-SAR missions in April and October 1994, we carried out near-simultaneous and intensive ground campaigns at our three test sites: Mammoth Mountain in the Sierra Nevada, California, the Urumqi River basin in the Tien Shan, China, and the Echaurren basin in the Andes, Chile. At Mammoth site, three corner reflectors were used for calibration. For each data-take we measured two snow pits to obtain vertical profile information at each test site. Snow property measurements in the profiles included temperature, density, grain size and size variation, free liquid water content, and stratification. In addition to snow pit data, we also measured two transects across each site to assess spatial distribution. Those measurements included snow depth, density, wetness, and surface roughness. An average value of five points measurements with one center point and four surrounding points at about 20 m apart was used to represent the snow depth and wetness values. At the Tien Shan site, five corner reflectors were used. Snow property measurements includes snow temperature, depth, and density. At the Echaurren site, we measured snow temerature, depth, density, particle size, and surface roughness. Those field measurements have been geo-located and integrated into our data-base and can be obtained through our web site at http://www.icess.ucsb.edu/hydro/hydro.html for other interested research groups.

II. Understanding the Links between SAR Measurements and Snow Parameters

Modeling Microwave Backscattering Response to Snow Covered Terrain

The key issue in estimating snow properties from remote sensing data is understanding the links between the electromagnetic interactions in different parts of the spectrum and the physical snow properties. Modeling microwave backscattering from snow packs requires knowledge of snow microstructure and characteristics. The input snow parameters for discrete dense media model include snow temperature, density, mean particle size, size variation and distribution function, stickyness parameter, and pair-distribution function. Usually, these parameters, except snow temperature and density, are based on many assumptions, and commonly used measurements such as grain size are not sufficient to describe the relationships with the observed microwave signals. To overcome this problem, we have developed a technique to measure ice particle size variation in addition to snow density and ice particle size by using the stereological data and have found that the ice particle size distribution in seasonal snow can be characterized as a log-normal distribution function (Shi, et. al., 1993). The required parameters (the geometric mean diameter and standard deviation of particle diameters) for fully describing the particle size variation and distribution can be directly measured by stereological variables. The optically equivalent ice particle size for Rayleigh scattering in a snowpack with grain size variations can be determined from the snow sections. More recently, we have developed another technique to determine the pair-distribution function from real snow section measurements (Zurk, et. al., 1997) so the all input snow parameters can be obtained from real snow measurements for the discrete dense media modeling.

In modeling the wave scattering from snow covered terrain, we expect a combination of surface and volume scattering. This is the case as snow has both irregular air-snow and snow-ground interfaces. For the surface and volume interaction components, the most common used method is to calculate them independently. We have developed a polarimetric model that includes both surface and volume scattering as well as the interaction terms between surface and volume (Shi and Dozier, 1993, Shi and Dozier, 1995). Radiative transfer model for dense media is used for the volume scattering component. The surface scattering model, IEM, is used to evaluate the surface scattering components and introduced to the radiative transfer equations as the boundary conditions in order to evaluate the importance of the interactions between the surface and volume scattering signals. We found that the surface and volume interaction terms are the important scattering source for cross polarized signals. The surface-volume interaction terms under the independent assumption results in an over-estimation for HH polarization. The error increases as the surface roughness increases, which can be as large as by a factor about 2 for a very rough surface. For VV polarization, however, it always over-estimates at small incidence angle and under-estimates at large incidence angle.

Backscattering Response to Snow Wetness at C-band

At C-band, the backscattering is controlled by snow volume backscattering and the surface backscattering at air - snow interface. When wetness is low, the dielectric contrast between air and snow is small and volume scattering dominates, so backscattering is not sensitive to surface roughness. As snow wetness further increases, backscattering becomes sensitive to surface roughness. This is because the surface scattering component becomes dominating, resulted from rapidly increasing surface scattering component and decreasing volume scattering component.

The relationship between backscattering and snow wetness is controlled by the scattering mechanism. When the surface is smooth, volume scattering dominates. Hence as snow wetness increases, both the volume scattering albedo and the transmission coefficients greatly decrease, resulting in a negative correlation between the backscattering and snow wetness. When the surface is rough, increasing snow wetness causes surface scattering to dominate, resulting in a positive correlation between the backscattering and snow wetness. Moreover, the relationship between co-polarization and snow wetness can be either positive or negative, depending on snow characteristics and surface roughness and on incidence angle.

Backscattering Response to Dry Snow Density at L-band

At long wave length (L-band with 24 cm wave length) snow particle size has little effect on the backscattering signals from a dry snow cover. The scattering mechanism can be considered as a homogeneous dielectric layer (snow pack) over a rough surface. When we simulate the backscattering coefficients at L-band for a bare surface with and without dry snow cover, the same ground dielectric and roughness parameters were used in the simulation, the backscattering will be much higher for snow covered condition, especially at large incidence angle. This is because when radar signal passes through a dry snowpack, there are several changes playing important roles in comparison to the interaction with a bare surface. First, it will result in a change of wave propagation constant because snowpack is a dense media. In other words, depending on snow density the wave length will shifted shorter so that the snow-ground interface becomes rougher. Secondly, it will also cause a change in incidence angle according to Snell's law. Thirdly, the dielectric contrast at snow-ground interface will be different with even the same ground without snow cover. The first factor is especially important at low frequency, such as L-band. At high frequency, however, the first factor becomes less important since the surface backscattering at snow-ground interface become independent of frequency according to Geometric Optics Model. Above scattering mechanism provides a great potential for developing an algorithm of inferring snow density from L-band SAR measurements.

Backscattering Response to Snow Depth at C- and X-band

The problem with estimation of dry snow depth by using active microwave sensors is that the backscattering power received from the radar depend not only on the total snow mass but also on snow density, grain size, structure factor, stratification and ground surface roughness and dielectric properties. Perhaps confusion resulted from a different combinations of above factors.

Through SAR data analyses and model simulations, we found that backscattering measurements from dry snow are affected by three sets of parameters: (1) sensor parameters, (2) snowpack parameters, and (3) ground parameters. The relationships between backscattering signals and snow water equivalence can be either positive or negative depending on the snow physical parameters, ground surface parameters and incidence angle. In addition to snow density and ice particle size, size variation, snowpack stratification, and underlying ground conditions affect the interpretation of the observed backscattering signals. When the scattering signal from the snowpack is greater than the signal from the ground (attenuated by the overlying snow), a positive correlation is expected. Otherwise, the correlation is negative. The key issue in development of an algorithm for estimation of snow depth is how to decompose or separate the ground backscattering signal with the snow volume backscattering signal in the SAR measurements.

III. Snow Mapping from Synthetic Aperture Radar

Through this project, we have developed several techniques for mapping snow in alpine regions.

• single-polarization technique: using the normalized C-band, single-polarization backscattering measurements by a function of incidence angle and also used texture information to discriminate wet snow cover and glacier ice from other targets (Shi and Dozier, 1993). The method requires a high-quality Digital Elevation Model (DEM), accurate knowledge of the sensor's flight path, and precise modeling of the angular backscattering behavior of both snow-free and snow-covered terrain.
• multi-polarization technique: using measurements of the polarization properties from multi-polarization C-band SAR to map wet snow and glacier ice without a DEM (Shi et. al., 1994).

However, both methods can map only wet snow, because it is difficult to discriminate dry snow from bare ground and short vegetation. In addition, the test areas in these papers were about 10 km 10 km and had only a few different surface covers within the scenes. For larger drainage basins-up to 100 km 100 km-there are likely many different targets within a scene, with consequently large variability in the backscattering and polarization signatures. Hence the above-described techniques might not be reliable. For example, backscattering from wet snow is similar to that from smooth dry soil, alluvial surfaces, and rough water under windy conditions. Similarly, the change-detection approach is less reliable because other natural environmental changes occur in the other surface covers.

In the study (Shi and Dozier, 1997), we evaluated the characteristics of the backscattering, polarization, and frequency ratios of the targets of our study site. The larger variations in backscattering and polarization property measurements are caused by the great variability of natural targets at SIR­C/X­SAR coverage scale. These make it difficult to map snow, even wet snow, with a single-frequency, single-polarization SAR. In our study site, wet snow could not be discriminated from smooth bare surfaces at C- or X-band, because there is not much difference in dielectric properties and roughness between these two targets. They can be separated at L-band because of the effects of flow structures in the snow. Dry snow is difficult to separate from short vegetation with typical backscattering measurements, but the contrast between dry snow and short vegetation at high elevation can be enhanced by using polarization filters.

We have developed two type of classifiers based on classification trees. The first type of the classifier was developed by using intensity measurements, polarization properties, and frequency ratios. It can map dry snow and discriminate dry from wet snow, but it requires topographic information for radiometric terrain correction and to reduce effects of local incidence angle. It is about 79% as accurate as a TM binary classification (as shown at top and bottom on right column of figure 1). The second type of classifier was developed based on polarization properties and backscattering ratios between different frequencies. Since these measurements can be obtained correctly without radiometric terrain calibration, the classifier does not require topographic information and can be used to map wet snow (as shown in left column of figure 1. From top to bottom, the classification results are from SAR, TM binary, and TM percent, respectively). Its accuracy is 77 % when compared with TM binary classification. In comparing our schemes to the snow-fractions derived from Landsat TM data, we found that both SAR and TM binary classification results underestimate total snow cover. The major difference between SIR­C/X­SAR and TM classification results is for pixels only partially covered by snow, especially in forested regions. We found that the dry and wet snow maps derived from SIR­C/X­SAR corresponded to the pixels with more than about 80% snow cover.

More recently, we evaluated the coherence measurements between two repeat-pass from first and second mission. We found that this measurement are are very low for both dry and wet snow covers between a snow covered scene and one without snow and provide a very good separation between snow cover and bare surface as well as short vegetation. These two targets are most difficult to discriminate with snow. Our current understanding and interpretation as follows: If the ground is completely undisturbed between viewings the signals will be highly correlated. In the case of bare surface, a change of soil moisture will result in a decorrelation. However, the amount of decorrelation is expected to be relative small because radar senses a same target with a same scattering mechanism (only change in magnitude). When measuring correlation between wet snow-cover and bare ground passes, the radar echoes will be close to completely decorrelated. This is because the radar signal in snow-covered pass can only penetrate a few centimeters so that the radar senses two different targets. For dry snow case, even the dominant scattering is from the interface of snow-ground, in addition to the change of dielectric properties, existing dry snow cover will result in large decorrelation due to change in local incidence angle when radar signal passes through snow layer which will cause a spatial baseline decorrelation. Thus, the correlation measurement provides a significant information to map snow covered area. We could use the coherence measurements to discriminate forest, open water, and snow with short vegetation and bare ground. Then using backscattering intensities and their changes to separate forest and open water with snow. Figure 1 in middle row of right shows such a classification.

VI. Inferring Snow Wetness Using C-band Image Data

Monitoring spatial and temporal changes of liquid water content in a snow pack is important for hydrological modeling because it identifies that a particular area of the basin can contribute immediately to runoff. Despite its importance, hydrologic models have not used snow wetness in analysis or forecasting of snow melt because the data have not been routinely available.

Our algorithm (Shi and Dozier, 1995) is based on a first-order scattering model that considers both surface and volume scattering. A simplified surface backscattering model was obtained from the numerical simulations for the conditions of most seasonal wet snow covers. Through this simplified surface backscattering model and the property of the volume scattering ratio in co-polarizations, which is only a function of snow permittivity, we have developed an algorithm for snow wetness retrieval using C-band polarimetric SAR imagery. Using three measurements , , and , we can minimized the effects of the snow volume scattering albedo and the surface roughness in order to estimate snow permittivity which can be directly related to snow wetness. The algorithm is applicable to the situations of incidence angle from to , and the snow surface roughness - rms height < 0.7 cm and correlation length < 25 cm. Figure 2 (left) shows the estimated snow wetness map from the data-take 40.4 on April 11, 1994. At regional scale, the inferred spatial distribution of snow wetness from the SIR-C data showed wetter snow on the south slopes than the north slopes, and wetter snow at lower elevations. The comparison of SAR-derived snow wetness with the ground measurements indicated that the absolute error at 95 percent confidence interval was 2.5 percent by volume. The inversion algorithm performs well using C-band SIR-C/X-SAR data and would prove useful for routine and large-area snow wetness measurements when a spaceborne multi-parameter SAR becomes available.

V. Estimation of Snow Density using L-band Measurements

Dry snowpack has great effect on backscattering coefficients even a little contribution of volume scattering from snowpack at L-band. The major scattering mechanism is due to the surface backscattering at snow-ground interface. The existing dry snow cover results in a wavelength shift, incidence angle change, and the dielectric contrast change when comparing without snow covered bare surface. At L-band, the effects of dry snowpack's volume scattering and extinction can be ignored. It has been understood that the co-polarized signature of surface scattering are highly correlated and sensitive to the local incidence angle, in addition to the dielectric and roughness properties of the surface. Based on the scattering mechanisms described above, we have developed a physical based inversion model (Shi and Dozier, 1996) for estimation of snow density using L-band three co-polarization measurements: VV, HH and correlation between them. This algorithm was based on the simulated backscattering for most bare surface condition by IEM model and the relation among those three measurements are very sensitive to the local incidence angle. By including effects of snow density in this relation, we can estimate the difference between the local incidence angle at air-snow interface (calculated by using DEM and the shuttle location) and the incidence angle at snow-ground interface (refractive angle in snow pack), which are directly related to the dielectric constant of snow pack by Snell's law and so to snow density. With estimated snow density, we can, then, estimate under ground dielectric constant and the surface roughness parameter using the algorithms, which are summarized in the last session, we have developed for estimation bare surface soil moisture and roughness parameter. Figure 2 (top right) shows an estimated snow density obtained from the data-take 67.1 on April 13, 1994.

VI. Estimation of Snow Depth using C- and X-band Measurements

Our inverse algorithm for estimation of snow depth (Shi and Dozier, 1996), is based on the first-order backscattering model. With estimated snow density, under ground dielectric and roughness parameters from L-band measurements, there are only three unknowns - snow volume scattering albedo and snow depth and ground surface backscattering coefficient. We have developed two models for the relations between C-band HH and VV and between C-band-HH and X-VV for ground surface backscattering components, based on the IEM model simulation. With those models, three unknown ground surface backscattering components at C-band HH, VV and X-band VV can be reduced to one. Therefore, with C-band VV, HH, and X-band VV backscattering measurements, we can estimate snow depth, snow pack volume scattering albedo, and ground surface backscattering component. Figure 3 (bottom right) shows the estimated snow depth in D and equivalent particle size in E maps obtained from the data-take 67.1 on April 13, 1994. Currently, we have finished the intensive test for each step involved in this algorithm and are preparing the manuscript for publication.

VII. Estimation Bare Surface Soil Moisture and Roughness Parameter

An algorithm based on the regression analysis of the simulated surface backscattering coefficients by the single scattering IEM model was developed to provide estimation of soil moisture and surface roughness parameter from L-band SAR co-polarized measurements over bare and short vegetated fields. Although the multiple scattering terms in the original expression for backscattering coefficients have been ignored, the single scattering IEM remained quite complex and its simplification was deemed necessary for practical applications. This simplification procedure presented in this paper took the form of a regression between the calculated backscattering coefficients and surface parameters (i.e., soil moisture, rms. roughness height, surface correlation function and correlation length). The results were equations that would relate the backscattering coefficients directly with surface parameters through some constants depending only on the incidence angle. The inverse models were checked against the single scattering IEM for its applicability by comparing the values of calculated backscattering coefficients over a wide range of soil moisture and surface roughness conditions. By going through a series of sensitivity tests and a consideration of SAR calibration accuracy, it was found that the co-polarized backscattering coefficients and their combinations were the best measurements for input to the algorithm for the retrieval of surface parameters. It was also established that the absolute calibration accuracy of both AIRSAR and SIR-C was adequate for providing a reasonable estimation of soil moisture and roughness parameters.

This algorithm was applied to a series of AIRSAR and SIR-C measurements obtained over the Little Washita River watershed near Chickasha, Oklahoma in June 1992, and April and October 1994. The results from this effort were summarized in two different ways. First, the values of retrieved soil moisture and surface roughness parameter for a number of bare and short-vegetated fields were compared with those sampled on the ground in near concurrence with the AIRSAR and SIR-C measurements. A reasonable agreement was found between the retrieved and measured values. The rms. errors of the comparison were estimated to be 3.4% and 2.0 dB for soil moisture and surface roughness parameter, respectively. Next, the retrieved surface parameters on a regional scale were examined for their temporal changes. Two dry-down sequences, one derived from AIRSAR measurements in June 1992 and the other from SIR-C measurements in April 1994 (in figure 3), showed the expected changes in soil moisture in sparsely vegetated areas. The retrieved surface roughness parameter remained relatively constant with time in these areas as expected. However, the algorithm either failed to provide a solution or gave unreliable estimation of surface parameters over regions associated with moderate or tall vegetation. The main advantage of the algorithm is that the regression analysis was performed on a large data set simulated for a wide range of surface parameters at very fine intervals, it is expected not to have the site-specific problem commonly associated with an empirical model derived from a limited observations.

References

J. Shi, R. E. Davis, and J. Dozier, "Stereological determination of dry snow parameters for discrete microwave modeling", Annals of Glaciology, pp. 295-299, vol 17, 1993.

J. Shi and J. Dozier, "Measurement of snow and glacier covered areas by single-polarization SAR", Annals of Glaciology, pp. 72-76, vol. 17, 1993.

J. van Zyl, B. D. Chapman, P. Dubois, and J. Shi, "The effect of topography on SAR calibration", IEEE Transactions on Geoscience and Remote Sensing, vol. 31, no. 5, pp. 1036-1043, 1993.

J. Shi, J. Dozier, and H. Rott, "Snow mapping in alpine regions with synthetic aperture radar", IEEE Transactions on Geoscience and Remote Sensing, vol. 32, no. 1, pp. 152-158, 1994.

J. Shi and J. Dozier, "Inferring snow wetness using SIR-C C-band polarimetric synthetic aperture radar", IEEE Transactions on Geoscience and Remote Sensing, vol. 33, no. 4, pp. 905-914, 1995.

M. M. Brugman, A. Pietroniro, and J. Shi, "Mapping Alpine Snow and Ice Using Remote Sensing at Wapta Icefield", Journal of Canadian Remote Sensing, vol. 22, no. 1, pp. 127-136, 1996.

L. Tsang, K. Pak, R. Weeks, J. Shi, and H. Rott, "Electromagnetic Wave Scattering from Real-life Rough-surface Profiles and Profiles Based on an Averaged Spectrum", Microwave and Optical Technology Letters, vol. 12, no. 5, pp 258-261, 1996.

J. Shi, J. Wang, A. Hsu, P. O'Neill, and E. T. Engman, "Estimation of bare surface soil moisture and surface roughness parameters using L-band SAR image data", In press, IEEE Transactions on Geoscience and Remote Sensing, 1997.

J. Shi and J. Dozier, "Capability of SIR-C/X-SAR Mapping Seasonal Snow Cover in Moutainous Areas", In press, Remote Sensing Environment, 1997.

J. Wang, E. T. Engman, P. O'Neill, J. Shi, and A. Hsu, "Evaluation of Current Soil Moisture Algorithms from SIR-C Image Data", In press, Remote Sensng Environment, 1997.

1. M. Zurk, L. Tsang, J. Shi, and R. E Davis, "Electromagnetic Scattering Calculated from Pair Distribution Functions Retrieved from Planar Snow Sections", In press, IEEE Transactions on Geoscience and Remote Sensing, 1997.

J. Shi, J. Dozier, and H. Rott, "Modeling and Observation of polarimetric SAR response to dry snow", Proceedings IGARSS '93, IEEE No. 93CH3294-6, vol. III, pp. 1042-1045, 1993.

J. Shi and J. Dozier, "Estimation of snow water equivalence using SIR-C/X-SAR", Proceedings IGARSS '96, IEEE No. 96CH35875, vol. IV, pp. 2002-2004, 1996.