INTERCOMPARISONS BETWEEN PASSIVE AND ACTIVE MICROWAVE REMOTE SENSING, AND HYDROLOGICAL MODELING FOR SOIL MOISTURE

Soil moisture estimates from a distributed hydrological model and two microwave remote sensors (Push Broom Microwave Radiometer and Synthetic Aperture Radar) were compared with the ground measurements collected during the MAC-HYDRO’90 experiment over a 7.4-km2 water- shed in central Pennsylvania. Various information, including rainfall, soil properties, land cover, topography and remote sensing imagery, were integrated and analyzed using an image integration technique. It is found that the hydrological model and both microwave sensors successfully pick up the temporal variation of soil moisture. Results also indicate the spatial soil moisture pattern can be remotely sensed within reasonable accuracy usingexisting algorithms. Watershed averaged soil moisture estimates from the hydrological model are wetter than remotely sensed data. It is difficult to conclude which instrument yield better performance for the studied case. The choice will be based on the intended applications and information that is available.


INTRODUCTION
Knowledge of soil moisture distribution in space and time is of considerable importance for many hydrological and agricultural applications. As a result of the inhomogeneity of soil properties, topography, land cover, and precipitation, soil moisture is highly variable both spatially and temporally. Soil moisture estimation has been based on extrapolation of point measurements.
Recent advances in microwave remote sensing have demonstrated the ability to measure surface soil moisture, in the order of 5 cm, under a variety of topographic and land cover conditions (Engman, 1990).
Despite the promising perspective of this new technique, its application to agricultural and hydrological sciences has been slow. This is because most existing hydrological models are formulated on point processes. These models are not capable of using the remotely sensed data as direct input or to verify output. This paper compares remotely sensed and model simulated soil moisture with ground observations using the data collected in an experiment conducted in the summer of 1990 (MAC-HYDRO'90). The spatially-distributed hydrological model proposed by  is used for simulations. The purpose of this study is to evaluate the performance of the hydrological model and to examine the limitations of various remote sensing techniques used in soil moisture estimation. Results will be used to investigate future directions for incorporation of the remotely sensed data into hydrological models. Site Description MAC-HYDRO'90 was conducted over a portion of the Mahantango Creek which is a 7.4-km 2 research watershed operated by the Northeast Watershed Research Center of the USDA, ARS in Pennsylvania. The climate of this area is considered as temperate and humid. The average annual precipitation and evapotranspiration for the watershed are 1128 mm and 479 mm per year, respectively. The intensive study area includes a subwatershed (WD38) of about 50 ha on the eastern portion of the larger area (see Figure 1). The 50-ha subwatershed is nearly all cropped (corn, wheat, oats, and hay) and is bounded on the south by forest. The soils within this region are primarily silt barns and barns, and contain 0.5 2.0 % organic carbon. Within the studied watershed, 15 different soil types can be identified. These soils have similar hydraulic properties (Troch et al., 1992).

Weather Conditions
The weather conditions for the experiment were dry initially. No rain was recorded during the preceding 5 days, resulting in uniformly dry soil conditions. After the first flight (July 10, 1990), there was an approximately 52 mm of rainfall over a four-day period, followed by a strong dry down. These conditions generated a wide range of soil moisture conditions which provide an excellent test ground for remote sensors and allow for intercomparisons among various data. The rainfall record and the dates of data collections are tabulated in Table 1.

Ground Data
Two kinds of soil sampling strategies were used. For large homogeneous agricultural fields, samples were taken from a grid to provide a field averaged soil moisture value. In addition, samples were collected along transects which were aligned at right angles to the streams. Samples were taken at two depths, 0 5 cm and 5 -~10 cm and consisted of 5 cm3 in volume. The location of some sampling sites and raingage network is shown in Figure 1.
Land cover information was compiled for a large area and was classified into 9 categories (see Figure 2). Rainfall records were collected from a network of 15 tipping-bucket raingages deployed over the watershed. A micrometeorological station located near the center of the catchment provide the time series of meteorological variables.

Passive Microwave Radiometer
The passive microwave instrument used in this campaign was the push broom microwave radiometer (PBMR). The PBMR operates at L-band (f = 1.42 GHz). It has four horizontally polarized beams pointing at ±8°and ±24°from nadir. The field of view is 1.2 times the altitude which was 300 m in MACHYDRO'90. For a detailed description of the PBMR, see Schmugge et al. (1988).
Data collected from the PBMR were processed following procedures that have been successfully employed in previous experiments (Schmugge et al., 1992). Vegetation corrections are applied to the average brightness temperature maps to estimate soil emissivity (Jackson and Schrnugge, 1991). Dielectric constants of soils are calculated from soil emissivity using Fresnel's formulae. Knowing the soil dielectric constants, a semi-empirical dielectric mixing model (Dobson et al., 1985) is used to estimate the volumetric soil moisture.

Synthetic Aperture Radar
Aircraft radar data were acquired at multiple angles over the Mahantango Creek using the Jet Propulsion Laboratory multipolarization imaging radar (AIRSAR) in three frequencies (1 = 0.44, 1.25 and 5.33 GHz). For a detailed description of the instrument, see Held et al. (1988). Three flight lines were flown each day with the objective of obtaining various incidence angles (20°,30°and 45°)of the target area (76°35'W, 40°43'N). On July 15 and 17, high resolution data with a 3.331 m slant range pixel size were also taken.
The AIRSAR imagery were calibrated for phase, cross-talk, channel imbalance and absolute power using trihedral corner reflectors. The underlying theories and algorithms for signal calibrations are presented in van Zyl et al. (1990). The calibrated SAR imagery are then registered with the USGS 7.5-mm digital elevation model (DEM), giving the local incidence angle of each pixel.
Hydrological Model The hydrological model predicts patterns of soil saturation and their relationship to both saturation excess and infiltration excess surface runoff generation by solving the three-dimensional Richards equation numerically . Richards equation with pressure head as the dependent variable can be written as where i is time, z is the vertical coordinate, positive upward, and the hydraulic conductivity is expressed as a product of the conductivity at saturation, K,, and the relative conductivity, Kr. An extension of the van Genuchten characteristic equations (van Genuchten and Nielsen, 1985) is used to describe the nonlinear relationships of volumetric moisture content 0, specific moisture capacity 5, relative hydraulic conductivity K, and the pressure head. Notice that hysteresis effects on moisture redistribution are not taken into account.
The initial water table depth for each pixel is computed using the procedure developed by Troch et al. (1992). The lower and lateral boundaries are assumed impervious. According to the geological records, the location of lower boundary is held fixed at 5 m below surface.

RESULTS AND DISCUSSION
Some large agricultural fields were used as verification sites to test the performance of instruments before the comparisons are performed.

Verification Sites
Four corn fields located east of the main watershed are chosen for verification purposes (see Figure  2). Data collected over these fields are also used to develop inversion algorithm for SAR. These corn fields are the largest accessible agricultural fields in the area. During the experiment, corn stood approximately 90 cm in height and contained 2 kg/rn 2 of water.   Figure 3 displays the temporal variation of the PBMR brightness temperature and the L-band HH-polarization SAR signal averaged over the corn fields 1 and 2 during the course of the experiment. Volumetric soil moisture contents from ground measurements are also plotted in the figure for references. It can be seen from the figure that the brightness temperatures measured by the PBMR decrease with increasing soil wetness. Meanwhile, stronger SAR backscattering signal was observed on wet days. In general, both sensors have reflected the temporal variation of soil moisture on these large corn fields pretty well.
Most existing SAR inversion algorithms are designed for bare soil surfaces (Soares et al., 1991;Oh et al., 1992). Pultz et al. (1990) have presented an estimation scheme for wheat and canola using field data collected in Canada. However, as pointed out by the authors, those relationships are site specific. It is, therefore, decided to develop empirical relationships for the MAC-HYDRO'90 site. Signals from four corn fields, two oat field and three pasture areas were extracted and linearly regressed with corresponding 0-~5cm ground soil moisture measurements. Results of the regression analysis are summarized in Table 2.
It appears that no particular combination of wavelength and polarization yield decisive edge. Considering the fact that it is more difficult to calibrate cross-polarization signal than like-polarization signal, we have decided to use the C-band VV-polarization signal to estimate soil moisture for all corn fields in watershed. For pasture and oat fields, the L-band VV-polarization signal will be used. These estimated regression relationships are shown in Figure 4. It is noted that the ranges of validity of these empirical relationships are limited. Extrapolation of the regression equations could lead to significant errors.

Subwatershed
To estimate watershed soil moisture from the PBMR brightness temperature, we apply the vegetation correction over the area in four categories: corn (38%), small grains (28%), pasture (14%) and hay (13%). Forest (6%) and residential area (1%) are excluded from the computation because the microwave signals are not related to soil moisture under these situations. The vegetation biomass for each category has been estimated from field samples or from previous data. where M~is volumetric soil moisture content in %~~and~are the VV-polarization backscattering coefficients in dB for C-band and L-band, respectively. It should be noted that, in order to reduce signal noises, field averaged signals are used in the above relationships.
The average soil moisture values over the WD38 subwatershed derived from the PBMR, the SAR and the hydrological model are listed in Table 3. The ground observations were averages of approximately 60 ground samples except on July 15 when only 33 were taken. The estimates between the PBMR and the ground measurements are in good agreement. This implies that the PBMR average procedure is quite successful in this case. Estimates from the hydrological model are wetter than other observations. The temporal variation, however, is correct. The cause for this bias is currently under study. Despite using the rather crude empirical relationships, the SAlt is able to predict watershed averaged soil moisture values within 20 % of the ground measurements.
Finally, it is difficult to compare the performances of passive and active microwave instruments under the current circumstance. The fine resolution of the SAR was partly diminished when field averaged soil moisture. In addition, The SAR requires additional topographic information than the PBMR. On the other hand, geo-referencing of the PBMR measurements is an involved work and is subject to large uncertainties for a small agricultural watershed, especially for a small watershed such as Mahantango Creek. The decision of which instrument should be used should depend on available information, as well as the data resolution required for the intended applications.

SUMMARY
The intercomparisons between hydrological model and microwave sensors were conducted over a small watershed in central Pennsylvania. Results can be summarized as follow, (1) The temporal variation of soil moisture patterns over the verification sites was successfully picked up by both passive and active microwave sensors.
(2) Both microwave instruments yield soil moisture estimates within 20 % of the ground measurements. Soil moisture estimates from the hydrological model are wetter than observations during the MAC-HYDRO'90 period. The choice of an appropriate instrument will depend on the intended applications and available information.    Figure 3: Temporal variation of brightness temperature and the L-band RH-polarization backscattering coefficient averaged over corn fields 1 and 2 during the course of the MACHYDRO'90 experiment. The local incidence angle of the SAR over corn fields 1 and 2 is approximately 39°.