Remote Sensing of High-Latitude Wetlands Using Polarized Wide Angle Imagery

Guillaume L. Perry^a, Joel A. Stearn^a, Vern C. Vanderbilt^a, Susan L. Ustin^b, Martha C. Diaz Barrios^b, Leslie A. Morrissey^c and Gerald P. Livingston^c, Francois-Marie Breon^d, Sophie Bouffies^d, Marc M. Leroy^e, Maurice Herman^f and Jean-Yves Balois^f

^aNASA Ames Research Center, 242-4, Moffett Field CA 94035 ^bUniversity of California, Davis, California, USA ^cUniversity of Vermont, Burlington, Vermont, USA ^dCEA/DSM/LMCE, Gif sur Yvette, France ^eUMR CNES-CNRS-UPS, Toulouse, France ^fL.O.A., U.S.T. de Lille, France

Further author information

Author	Email	Telephone	Fax
Guillaume L. Perry	gperry@gaia.arc.nasa.gov	415-604-0269	415-604-4680
Vern C. Vanderbilt	vvanderbilt@mail.arc.nasa.gov	415-604-4254	415-604-4680
Joel A. Stearn	jstearn@gaia.arc.nasa.gov	415-604-0636	415-604-4680

Abstract

Representing the areal extent of circumpolar wetlands is a critical step to quantifying the emission of methane, an important greenhouse gas. Present estimates of the areal extent of these wetlands differ nearly seven fold, implying large uncertainties exist in the prediction of circumpolar methane emission rates. Our objective is to use multi directional and polarization measurements provided by the French POLDER sensor to improve this estimate. The results show that wetlands can be detected, classified and their area quantified using the unique, highly polarized angular signature of the sunglint measured over their water surfaces.

1. Circumpolar Wetlands: A Major Source of Methane

Circumpolar wetlands are a major source of atmospheric methane. As a greenhouse gas, methane plays an important role in Earth's radiative budget. In order to improve prediction of climate change, one needs to know precisely what are sources and sinks of methane, their areal extent and their total emission rate. Present estimates of the areal extent of circumpolar wetlands differ nearly seven fold, providing a large source of uncertainty in methane budgets^1,2,3. In order to improve this situation, better satellite monitoring of wetlands is needed.

In wetlands, methane is produced in inundated soil by an anaerobic microbial decomposition process. The production and emission of methane in inundated areas are controlled by different factors^4,5,6,7 such as availability of organic matter, presence of anaerobic conditions, temperature and the presence of vegetation. Due to such different production and emission processes, rates of methane exchange with the atmosphere can vary over two to three orders of magnitude among different wetland types, principally between open water areas (such as lakes) where there is no above-water vegetation and inundated herbaceous wetlands. Consequently, to successfully quantify the methane exchange rate, the analysis of remotely sensed data will have to fulfill two objectives of equal importance: quantification of the areal extent of wetlands and non wetlands and discrimination of vegetated wetlands from open water areas. Although optical remote sensing methods are successful in finding deep, clear open water areas they routinely fail in the detection of vegetated wetlands. Vegetated wetlands typically have spectral signatures indistinguishable from those of other community types which release small amounts of methane. No prior optical remote sensing analysis of the sunglint has retrieved information about wetlands and quantified their areal extent.

Here we demonstrate that the sunglint signature of wetlands may be used to discriminate wetlands and non wetlands and to estimate the areal extent of non inundated vegetation, inundated vegetation and open water areas. Multidirectional data acquired with POLDER over BOREAS show that the sunglint presents a unique polarized angular signature due to the physical principles of the specular reflection^8,9.

2. Hypotheses To Be Tested

We have tested three hypotheses: First, the sunglint reflected by water surfaces provides a strong, unique, angular, highly polarized signature reflection characteristic of inundated, CH₄-producing areas and uncharacteristic of non-inundated, non-CH₄-producing cover types. Second, inundated wetlands and open water areas display different, wind dependent glitter signatures as a function of view angle. Third, as a consequence, the analysis of remotely sensed data, collected in and near the specular direction, will allow highly accurate discrimination of CH₄ source areas, namely inundated wetlands (areas with high CH₄ exchange rates), open water (low CH₄ exchange rates) and non-inundated (non-CH₄-emitting) cover types.

3. Data Collection

Polarized radiance images were collected using POLDER (POLarization and Directionality of Earth Reflectance), a large field of view, multispectral, imaging sensor¹⁰ developed by the Laboratoire d'Optique Atmospherique (University des Sciences et Techniques de Lille Flandres Artois, Lille, France). This sensor has been mounted in a C-130 airplane flying approximately five km above the Southern Study Area (SSA) of the Boreas Experiment, located in the region near (54°N, 105°W) in central Canada not far from Prince Albert, Saskatchewan. The radiance imagery were acquired on two dates - June 1 and July 21, 1994 - in the solar principal plane above five test sites: Old Aspen, Old Black Spruce, Old Jack Pine, Young Jack Pine and Fen. On July 21 additional radiance images of the SSA were collected along numerous parallel flight lines, Fig. 1, each flown in the solar principal plane either toward or away from the solar azimuth direction. On both dates concurrent data of surface wind speed were collected at the Old Aspen and Old Jack Pine test sites¹¹. During acquisition of POLDER imagery, large format color infrared aerial photographs, each 23 cm x 23 cm, were collected from the C-130 aircraft using a high spatial resolution Zeiss mapping camera equipped with a 153 mm focal length lens. Photographs taken on the * flightline, Fig. 1, included the subsolar or specular direction.

During flight the POLDER images were collected periodically, at 10 second sampling intervals, in five spectral bands in the blue (443 nary), green (550 nary), red (665 nm) and near infrared (865 nm and 910 nm) spectral regions. Three crossed polarizers in the near infrared band (865 nm) allowed estimation of Stokes polarization parameters (I, Q. U) which were used to estimate the normalized radiance, the polarized portion of the normalized radiance and the angle of polarization. The field of view of each image was approximately ± 53° along track (in the principal plane) and ± 45° across track (in the perpendicular plane). Because of the large field of view of the POLDER sensor, each image included both the subsolar (or specular direction) and the antisolar (or hot spot direction). Between 92% and 94% of each image overlapped adjacent images. This implies that a ground point was observed in multiple images and consequently seen as a function of different viewing directions.

4. Analysis of Polder Data

POLDER data were calibrated with reference to sensor calibration measurements completed during May 1994 immediately prior to the start of the field campaign. The image data, collected as digital numbers proportional to scene radiance, were converted to normalized radiance values by multiplying by p and then dividing by the value of the solar irradiance at the top of the atmosphere. To improve image band-to-band registration accuracy, the spatial resolution of each image was degraded by averaging 3x3 pixel regions, increasing the size of the ground pixel footprint from 50m to approximately 150m. Data were displayed, Fig. 2, using a Lambert conformal projection, removing the effects of changing sensor altitude and aircraft roll, pitch and yaw. The pixel data of all POLDER imagery collected within a flight line were redistributed to create fifteen new images, each representing a view angle ± approximately 3° in the principal plane along the flight line, Fig. 2. The results, presented in Fig. 2, show both the normalized radiance and the polarized portion of the normalized radiance. (Henceforth, these will be termed radiance and polarized radiance, respectively).

4.1 Analysis of radiance measurement

Fig. 2 shows sunglint near the specular direction (~ 42°) over areas covered with water, especially the lake and the large area of inundated vegetation represented by the white areas near the center line of the gray scale images. Fig. 2 shows also the wide difference between the angular width of sunglint observed above the inundated vegetation and the width observed above the lake (open water). The results, Fig. 3, show the radiance observed in two directions of observation. On the X axis we have reported the radiance measured near the specular direction. The Y axis corresponds to a direction 6 away from the specular direction. In this scatterplot, Fig. 3, each point represents a pixel observed in the principal plane (~ middle line of one segment). We identified three classes in this scatterplot of directional space and then established, using photointerpretation techniques, a correspondence with the following information classes:

1. The vegetation class is characterized by a group of pixels having low values in the two directions.

2. Due to the specular reflection and the ruffled nature of the open water surface, the pixels representing open water display large radiances in both directions and cluster in the upper right of Fig. 3.

3. The inundated vegetation, represented in the diagram by a line of pixels parallel to the X-axis, reflects light mainly in the specular direction and little in a direction 6 from the specular direction.

This shows that the difference in the angular width of the glitter is important and allows discrimination between types of wetlands, especially inundated vegetation and open water.

The results, Fig. 4, display the per class means of the data of non inundated vegetation, open water and inundated vegetation presented in Fig. 3.

4.2 Analysis of polarization measurement:

Repeating the radiance analysis (Section 4.1) but using polarized radiance, we find, Fig. 2, that lakes and inundated vegetation areas appear very bright and that non inundated vegetation features disappear, displaying a very low degree of polarization compared to that observed on inundated areas. Consequently, the scatterplot representing the distribution of polarized radiance (

) measured in the two directions (specular direction, 6° away from the specular direction) shows a similar separation between the three classes (Non inundated vegetation, inundated vegetation, open water). Fig. 4 shows the polarized signature of these three classes.

4.3 Classification of POLDER data and validation:

Using the two directions described in the section 4.1 (see also scatterplot in Fig. 3), a simple classification has been performed using only thresholds to distinguish open water, inundated vegetation and non inundated vegetation. Fig. 5 shows the classification map, 150 m (one pixel) wide extending 47 km along the principal plane, the centerline of the '*' flightline, Fig. 1. All pixels classified as open water represent the lake (pixel numbers 40-75). Close inspection of 11 of the color infrared, large format airphotos, collected concurrently with POLDER imagery, revealed that no other open water areas appeared within the one pixel wide flightline. Comparison between airphotos and classification results suggests that pixels (numbers 25-40) classified as inundated vegetation represent primarily inundated black spruce vegetation and are thus correctly classified. But in several instances the pixels classified as inundated vegetation represent clear cut areas containing small (10 m to 20 m across) open water areas. And interpretation of the air photos showed that pixels representing several rivers and streams, located in the vicinity of pixel number 260, were classified as inundated vegetation. Table I presents the results of the classification.

4.4 Preparation of the unmixing analysis:

Unlike the spatial resolution (150 m) of the airborne POLDER data analyzed here, the spatial resolution (~6 km) of the satellite borne POLDER will be significantly lower. Each satellite pixel will usually represent more than one land cover type; the pixel radiance and Stokes parameters will be a linear combination of different endmember signatures: R_ow, R_iv, R_niv, Q_ow, etc.:

where ow = open water, iv = inundated vegetation, ni = non inundated vegetation and a represents the proportion of pixel occupied by a specific cover type.

Q_ni and U_ni are near 0 because, as we have seen previously, the degree of linear polarization above vegetation is extremely small compared to that of inundated areas.

To obtain the bidirectional light scattering properties of a wetland region as it would be observed on a mixed pixel, we have assumed that all landcovers (open water, inundated vegetation, non inundated vegetation) are randomly distributed on the ground. Let us consider one POLDER image. Each pixel of one image can be associated with one angle and with one geographical position that can be covered with open water, inundated vegetation or non inundated vegetation. Averaging the images of a flightline for each direction of observation, we obtain a 'composite signature' linearly related to the mean signature of each class:

where prob represents the probability of each class.

In consequence, at each angle of observation, by adding the multiple images of a flight line, we will obtain a composite image that will be a linear combination of the different signatures (open water, inundated vegetation and non inundated vegetation) observed in one segment. See also Fig. 6 .

From a satellite observation of a mixed radiance Rmes, we want to retrieve the relative proportion of open water, inundated vegetation and non inundated land covers. Knowing that the mixed radiance observed is a linear combination of the three mean signatures, we have to invert the system:

for each direction #i available.

We inverted the mean (or composite) radiance, Fig. 6, of a flightline in order to obtain the coefficients a , Table I. Fig. 7 shows the composite image in the principal plane and the normalized radiance obtained after linear combination of the three mean signatures.

4.5 Use of polarization in the unmixing analysis:

Fig. 8 shows the degree of linear polarization estimated after averaging images to obtain a composite image as described Section 4.4. We have chosen to plot only the directions in the principal plane (j _v = 0). Both the polarized radiance and the degree of polarization are only significant near the specular direction and tend to zero elsewhere, showing again that polarization due to vegetation is neglectible compared to that for wetlands observed in the specular direction.

To summarize we can express radiance and Stokes parameters for a mixed pixel as:

Using polarization simplifies our analysis since we do not depend on vegetation nor on the presence of sediments in open water.

5. Discussion

5.1 Validity of hypotheses:

1) The visually blinding sunglint provides a strongly polarized, directional signature characteristic of inundated areas and uncharacteristic of non inundated. This result is supported by Fig. 4 that demonstrates the radiance and/or polarized radiance of inundated vegetation and open water are much larger than that of vegetation alone and shows extremely large variation with view angle around the specular direction. Such an angular feature, while characteristic of planophyll plant canopies (canopies with perfectly horizontal leaves), is uncharacteristic of the non planophyll plant canopies found in the vegetated areas of the flightline.

2) Second Hypothesis: Inundated wetlands and open water areas display different, wind dependent glitter signatures as a function of view angle. This hypothesis, illustrated in Fig. 3, is supported by the results Fig. 2 For the intermediate wind speeds, 2 m/s to 4 m/s, measured during image data collection, the angular spread of the glitter from open water areas (~ 25° ) is much greater than that from inundated vegetation areas (~ 5° ). At these wind speeds, the wind sheltering effects of the emergent vegetation, which grows above the water surface, will allow separation of inundated wetlands and open water areas in remotely sensed data. However, when wind speed is zero and therefore the surfaces of both vegetated wetlands and open water areas are equally smooth, then the angular spread of the glitter of each would be identical, causing them to be inseparable in remotely sensed data. And presumably at sufficiently large wind speeds, the water surfaces of both vegetated wetlands and open water areas will likely be equally ruffled; the angular spread of the glitter will be equally large and vegetated wetlands and open water areas again would be inseparable in remotely sensed data. Thus, the second hypothesis appears correct for the intermediate range of wind speeds, 2 m/s to 4 m/s, measured during image data collection. Presumably the second hypothesis will fail as wind speed approaches zero or, conversely, as wind speed becomes sufficiently large.

3) Third Hypothesis: Analysis of remotely sensed data, collected in and near the specular direction, will allow highly accurate discrimination of inundated wetlands (areas with high CH₄ exchange rates), open water (low CH₄ exchange rates) and non-inundated (non-CH₄-producing) cover types. This hypothesis is supported by the results, Fig. 5, which appear to represent a significant improvement in accuracy when compared to a land cover map, developed for the Southern Study Area of the Boreas Experiment, having a classification accuracy for inundated vegetation estimated to be 11%¹².

Except for a river misclassified as an inundated wetland, the classification results, Fig. 5, appear to be in good agreement with a wetlands map of the flightline which we developed by applying photo interpretation methods to large format aerial photography acquired concurrently with the POLDER imagery. Misclassification of rivers as a vegetated wetland might be due to the wind sheltering effects of tall woody vegetation located next to the river surface. The presence or absence of glitter, blindingly bright and uniquely associated with water, facilitated the photointerpretation of the wetlands along the flightline. But when we excluded from the photointerpretation process that portion of the imagery representing the specular direction, we discovered that we often confused inundated vegetation and non inundated meadows because both appear to be green vegetation when observed in non specular directions.

5.2 Mixed pixels and source of variability:

The ground footprint of many pixels, an area approximately 150 m in diameter, presumably includes more than one cover type. These mixed pixels appear in the results, Fig. 3, at locations between the clusters of pure pixels, representing vegetation, inundated vegetation and open water and designated as "pure" because they appear to contain only one cover type. Pure pixels within the open water cluster, Fig. 3, appear somewhat dispersed rather than tightly clumped about one location, due presumably to differences in the interaction process between wind and the water surface. These differences may depend in part upon the geographical size and depth of the open water area and upon the sheltering provided by nearby vegetation and land forms.

The inundated vegetation line segment, terminated by pure pixel clusters representing, respectively, inundated vegetation and non inundated vegetation, presumably includes along its length mixed pixels representing varying proportions of these two cover types. In addition the inundated vegetation line may represent a second potential source of variation due to the light transmitting properties of the emergent vegetation in vegetated wetlands. The varying amounts of vegetation emergent above the water surface of inundated vegetation will reduce by varying amounts the magnitude of the sunlight, represented by a position along the inundated vegetation line, which successfully traverses the canopy twice - from the sun to the water surface and from the water surface to the sensor. Interception of little/most sunlight by the canopy suggests that a data point will be located presumably close to the inundated vegetation/vegetation pure pixels, respectively. This means that the ground footprint of some pixels located along the inundated vegetation line includes only vegetated wetlands, not a mixture of inundated vegetation areas and vegetation. In other words the inundated line presumably includes along its length some pure inundated vegetation pixels.

5.3 Introduction to the unmixing analysis:

Extraction of information from mixed pixels is critically important for data received from the satellite version of the POLDER sensor for which the pixel ground footprint varies between approximately 6 km and 10 km. The test of the unmixing analysis using the composite image obtained through the averaging process discussed in Section 4.4 demonstrates the feasibility of the method. The inversion of the system provides estimates in good agreement with those obtained through photointerpretation and through classification (See Table I).

5.4 Use of polarization:

Results, obtained through the analysis of polarization, show a similar sensitivity to the different types of wetlands and non wetlands. The results show also that pixels representing only vegetation are concentrated near the origin. This suggests that the use of polarization will increase the contrast between inundated and non inundated area. This result is also noticeable in the mean polarized signatures of non inundated vegetation, open water and inundated vegetation represented in Fig. 4. The results show that the polarized light estimated for non inundated vegetation is insignificant compared to that from inundated vegetation and open water. This allows polarized light from non inundated vegetation to be ignored, simplifying our future analysis of mixed satellite pixels.

Conclusion

The polarized sunglint and its observation with a wide field of view imaging sensor such as POLDER is potentially a powerful new approach for classifying and estimating the areal extent of northern, high latitude wetlands (specifically open water areas and inundated vegetation) and non wetlands. The use of polarization in addition to the multidirectional measurement of radiance will enhance our ability to separate the wetlands and will allow us to better characterize the vegetation covering inundated areas. This would consequently improve methane budgeting, the ultimate goal of our research.

References

1. I. Fung, J. John, J. Lerner, E. Matthews, M. Prather, L.P. Steele, and P.J. Fraser, "Three-dimensional model synthesis of the global methane cycle," Journal of Geophysical Research, vol. 96, pp. 13033-13066, 1991.

2. I. Aselmann and P. J. Crutzen, "Global distribution of natural freshwater wetlands and rice paddies, their net primary productivity, seasonality and possible methane emissions," J. Atmos. Chem., vol. 8, pp. 307-358, 1989.

3. E. Matthews and I. Fung, "Methane emission from natural wetlands: global distribution, area, and environmental characteristics of sources," Global Biogeochemical Cycles, vol. 1, pp. 61-86, 1987.

4. L.A. Morrissey, S.L. Durden, G.P. Livingston, J.A. Steam, and L.S. Guild, "Differentiating methane source areas in Arctic environments with multitemporal ERS-1 SAR data," IEEE Trans. Geosci. and Remote Sensing, vol. 34, pp. 667-673, 1996.

5. N.T. Roulet, A. Jano, C.A. Kelly, L.F. Klinger, T.R. Moore, R. Protz, J.A. Ritter, and W.R. Rouse, "Role of the Hudson Bay lowland as a source of atmospheric methane," Journal of Geophys. Res., vol. 99, pp. 1439-1454, 1994.

6. D.S. Bartlett, D.S., K.B. Bartlett, J.M. Hartman, R.C. Harriss, D.I. Sebacher, R Pelletier-Travis, D.D. Dow and D.P. Brannon, "Methane emissions from the Florida everglades: patterns of variability in a regional wetland ecosystem," Global Biogeochemical Cycles, vol. 3, pp. 363-374, 1989.

7. K.B. Bartlett, P.M. Crill, R.L. Sass, R.C. Harriss and N.B. Dise, "Methane emissions from tundra environments in the Yukon-Kuskokwin delta, Alaska," J. Geophys. Res., vol. 97, 16645-16660, 1992.

8. C. Cox and W. Munk, "Measurements of the roughness of the sea surface from photographs of the sun's glitter," J. Opt. Soc. Am., vol. 44, pp. 838-850, 1954.

9. F.M. Breon and P.Y. Deschamps, "Optical and physical parameter retrieval from POLDER measurements over the ocean using an analytical model," Remote Sensing Environment., vol. 43, pp. 1-20, 1992.

10. Deschamps P.Y., F.M. Breon, M. Leroy, A. Podaire, A. Bricaud, J.C. Buriez and G. Seze ''The POLDER mission: Instrument characteristics and scientific objectives" IEEE Trans., Geosci. Remote Sensing, vol. 32,. pp. 598-615, 1994.

11. S.R. Shewchuk, "The Surface Mesonet for Boreas," J. Geophys. Res., 1997, (in press).

12. F.G. Hall, D.E. Knapp and K.F. Huemmrich, "Physically-Based Classification and Satellite Mapping of Biophysical Characteristics in the Southern Boreal Forest," J. Geophys. Res., 1997, (in press).

1998, Center for Spatial Technologies and Remote Sensing (CSTARS)
University of California, Davis