Color aerial video imagery of the same wetland was obtained at two different phenological stages, summer and fall of 1994, and at varying altitudes. Frames were georeferenced, rectified, mosaicked, and composited into a multi-temporal image. Vegetation was classified using a supervised classification routine. Comparison of the classified video image with the Thiessen polygon maps showed a 60% correspondence. Although the maps generated using field data provided more detailed classification of plant community types, the airborne video maps provided a more detailed spatial depiction of their distribution.
Recent advances in GIS, GPS, interpolation techniques, multivariate statistics, and airborne remote sensing are providing ecologists with new tools for constructing spatial databases of vegetation. We classified and mapped wetland vegetation using multi- temporal airborne videography and multivariate statistical analysis of field data. A global positioning system (GPS) was used to georeference both data sets. Image analysis and GIS interpolation techniques were used to prepare maps from the georeferenced airborne videography and field data sets, respectively.
Sample points were established in transects 20 m apart, oriented perpendicular to the river using a compass and fiberglass tape measure. Each transect consisted of a sample point in the riverbed, on the levee, and at 2-5 locations in the backwater. Sample points were spaced ~20 m apart along the transects in backwater areas, but were closer (4-10 m) in the riverbed and on the levee, where environmental gradients changed rapidly. A total of 81 sample points were established, and marked with steel rods.
Georeferencing The location of each sample point was determined with real-time differential processing (Hurn 1989) using two Motorola LGT 1000 global positioning systems (GPS). One LGT 1000 was used as a portable base station by positioning it over a National Geodetic Survey (NGS) benchmark overlooking the Pokegama River valley. Latitude, longitude, and elevation specifications obtained from NGS for the benchmark were programmed into the base station GPS. This GPS was linked to a UHF radio, which transmitted instantaneous correction data every 5 seconds to the field unit. Power was provided by two 7 amp-hour 12VDC utility batteries, enough to run the GPS and radio transmitter for about 10 hours. The GPS, radio, and battery fit into a 10"x10"x10" wooden box for convenient carrying. The same base station equipment was used to provide real-time differential correction for the GPS-linked airborne video (see below).
The Motorola LGT 1000 GPS used as the field unit was connected to a hand-held UHF radio, which received the correction data and provided voice communication with the base station. Both pieces of equipment were mounted onto a back pack frame for portability. The detachable GPS antenna was mounted onto a 2 foot PVC pole, which was positioned over the steel rods marking each sample point, and which could be elevated above any tall herbaceous vegetation which might obscure satellite signal reception (Fig. 1). Readings were taken for approximately 1.5 minutes per location, yielding a minimum of 100 GPS readings that were averaged to obtain the final coordinates.
Vegetation and Environmental Field Data Plant cover by species was measured in August 1993 using the line intercept method (Mueller-Dombois and Ellenberg 1974). Cover at each sample point (referred to hereafter as a "plot") was recorded over one meter of line oriented parallel to the river, and over a second meter of line oriented perpendicular to the first, because of the potential for anisotropy. All intercepting species were recorded, but where two or more species overlapped, only the dominant (overstory) layer was measured, such that total cover never exceeded 100% of the line lengths measured. Taxonomic keys (Fassett 1957, Britton and Brown 1970) and pressed specimens from the Olga Lakela Herbarium at the University of Minnesota, Duluth were used in plant identification. Six environmental variables were also measured for each sample point: water depth, soil organic matter, sand content (0.05 to 2.0 mm particles), silt content (0.002 to 0.05 mm particles), clay content (< 0.002 mm particles), and soil pH.
Multivariate Statistical Analysis of Field Data There are two basic conceptual models for statistically analyzing matrices of ecological data (Pielou 1984). "Classification" arranges sample points into discrete groups, such as plant communities. "Ordination" arranges sample points and/or species along environmental gradients. We used examples of both to analyze the field data collected: two-way indicator species analysis (TWINSPAN) is a classification technique, and Canonical Correspondence Analysis (CCA) is an ordination technique. FORTRAN computer programs are available in the public domain for performing these analyses (Hill 1979, ter Braak 1991).
Two-way indicator species analysis (TWINSPAN) is a divisive classification technique devised to partition reciprocal averaging (RA) ordinations (Pielou 1984). It treats all plots as a single entity at the onset, and iteratively divides them into hierarchical groups. TWINSPAN performs a one-dimensional RA ordination and breaks the axis at the centroid so as to crudely divide the data points into two classes; this procedure is repeated with the species quantities weighted in such a way as to emphasize the influence of especially useful diagnostic species (i.e. "indicator" species) identified by the first ordination. TWINSPAN produces a dendrogram showing relationships between plot groupings, such that groups positioned close to each other in the dendrogram are most similar. Only vegetation data were used in the TWINSPAN analysis performed here.
Canonical Correspondence Analysis utilizes both vegetation and environmental data. CCA is a direct gradient analysis technique, in which species composition is directly and immediately related to measured environmental variables. It mathe-matically defines canonical axes, and produces diagrams of sample points and plant species plotted in relation to those axes. The Cartesian coordinates of each sample point relative to the first three canonical axes were used in subsequent GIS analysis. The mechanics and advantages of CCA over other ordination techniques are described in detail by Palmer (1993).
It should be stressed that neither TWINSPAN nor CCA utilize spatial statistics, so the location of each sample has no bearing on its statistical treatment. Each program mathematically synthesizes information about vegetation, and describes species assemblages at each point in space that was sampled. Also, neither program produces vegetation maps; the interpolation capabilities of a GIS are required to generate maps from these point data.
GIS Analyses The two multivariate statistical methods yielded different types of output, so different GIS methods were needed to construct maps from their results. TWINSPAN assigned each sample point to a discrete vegetation class, suitable for GIS analyses involving categorical (nominal) data. These data were used to generate Thiessen polygon maps (Green and Simpson 1978) with the ARC/INFO Geographic Information System. Boundaries between Thiessen polygons of like vegetation class were dissolved in ARC/INFO. A data layer of the wetland boundary was generated by displaying the sample points on the computer screen and circumscribing the boundary based on its appearance on the airborne video image (Fig. 2); this data layer was used to clip off Thiessen polygons that extended past the wetland boundaries.
CCA generates canonical axis values for each sample point. As with any ordination technique, these are unitless numbers which may be positive or negative. Large positive numbers indicate sample points with a strong positive relationship to the CCA axis, and large negative numbers indicate those with a strong negative relationship. Point data layers were constructed for each of the first three CCA axes by assigning canonical axis values to their corresponding sample point locations. These point values were then interpolated using an inverse squared distance weighted moving average (Burrough 1986) with the IDW command in ARC/INFO GRID, resulting in grid of cells. The interpolation was performed using the 12 closest sample points, with the wetland boundary used as a barrier to limit the search for input points. The output cell size was 5 x 5 meters. The grids were converted to ERDAS Imagine files, and these files were joined in Imagine using the "stacklayer" command to create a three-band image. Each of the three bands was assigned a different color (red, blue, green) for simultaneous display. In the resulting color composite map, different color combinations indicated different combinations of canonical axis values.
Airborne Videography Video imagery was acquired in July and September 1994 from a Cessna Skyhawk aircraft at flight altitudes of 2000 and 1000 feet, respectively. Flights took place between 10 am and 2 pm to minimize shadow length. On-site observations were made within two days of the actual flight to account for phenological variables such as plant growth stage and background characteristics such as turbidity and amount of open water. Styrofoam targets were placed in areas of homogeneous wetland vegetation for use in supervised training procedures during subsequent image processing and to provide ground control data for verifying the video georeferencing (Sersland et al., 1995). The ground location of each of the eight targets was determined by processing carrier phase GPS information collected by Motorola LGT1000 receivers.
Aerial video imagery was recorded using a Panasonic GP-KR412 VHS color composite video camera head with a 12 mm focal length. This camera senses in the visible portion of the electromagnetic spectrum, from 0.42 - 0.85 um, though the exact spectral response in the red, green, and blue has not been determined. The video signal was recorded on 1/2 inch video tape in a Panasonic AG-7355 S-VHS video tape recorder connected to the camera head.
A Motorola LGT1000 Global Positioning System (GPS), programmed to receive a locational fix once a second, was integrated with the aerial video camera system on the July flight. Real-time differential corrections were provided by another LGT1000 unit on the ground using a UHF radio link. The Motorola was replaced by a 24-channel Ashtech 3DF ADU receiver on the September flight. This receiver computed the plane's position twice a second, and determined its attitude (roll, pitch and yaw) from four GPS antennas mounted on each wing and the front and rear of the fuselage. A data encoding system was installed in the VCR that interfaces with the GPS receiver to frame register the video data.
The in-plane system was controlled by a notebook computer. For each GPS location, the current SMPTE code (unique identifier) was queried from the video recorder. The latency in the GPS data was adjusted for and the result was a data file containing plane position and attitude information for two video frames each second.
In the lab, a 486 PC with a Truevision Targa+ frame-grabber connected to the Panasonic AG-7355 VCR automatically positioned the video tape to a given frame by entering the frame's SMPTE code and then rasterizing the image. After the selected images were frame- grabbed, they were automatically rectified, converted to a common scale, and mosaicked together into one georeferenced file using software written for PC ERDAS 7.5. The June and September images were co-registered to make a multi-date composite, using the targets and other identifiable objects for ground control. The two images were then layered using the routine LAYERSTACK in Imagine to produce a six band image. Training signatures were defined in the Signiture Editor with ERDAS Imagine's Area of Interest/Seed Properties, and merged until the maximum separation between classes was achieved. A supervised classification with the maximum likelihood decision rule was applied. The vegetation field data collected in August 1993 (see above) were used to evaluate classification accuracy. Details of this portion of the study are described by Sersland et al. (1995).
Real-time differential readings using conventional (C/A) GPS codes provided accuracies consistent with the spatial scale of the study, within 2-3 m of true ground location at the 95% probability level. If we had not used differential processing, the GPS readings would have been accurate only to within 100 m at the 95% probability level. The use of real-time differential processing overcame a potential equipment limitation caused by the steep terrain of the Pokegama River valley: given that the 6-channel Motorola LGT1000 GPS units could each receive signals from only six satellites, there was the potential that the base unit could have been receiving signals from satellites blocked from the field unit's view by the steep valley sides. Real-time differential processing allowed us to make sure that the same satellite signals were being received by both the field and base units, and that the GPS data quality was adequate prior to leaving the field site.
The use of real-time differential processing also aided field and aerial navigation. A number of the steel rods marking field sample points were submerged by high water levels during the spring after they were installed, but we were able to retrieve them by using real- time differential GPS to navigate to the coordinates we had measured the previous field season. In the airplane used for video image acquisition, the coordinates of target sites were pre-programmed into the airborne GPS to serve as "waypoints" to guide navigation, providing the pilot with information about the distance and direction to the target areas. Without real-time differential correction, the GPS readings would have been too imprecise for navigation.
The Ashtech 3DF ADU is a more expensive and specialized GPS, designed to provide attitude data in real-time to accuracies of about 0.1 degrees with a 2.5 meter antenna separation. These attitude data were used to correct video image georeferencing errors induced by aircraft roll, pitch, and yaw (Fig. 2). Post-processed or real-time accuracies were within 5 meters for each position computed (up to twice a second).
Although field workers were directed to set up the field sampling locations on an evenly-spaced grid within the backwater area, the GPS data revealed a large gap in the sampling grid behind a convex curve of the levee (Fig. 2). If we had assumed that the grid was rectilinear rather than determining GPS locations for each of the sample points, the vegetation maps resulting from the field data would have been distorted in this region. GPS georeferencing also allowed us to interface the field maps with other georeferenced data, such as National Wetland Inventory maps, which would not have been possible using coordinates determined by tape measure.
TWINSPAN/Thiessen Polygon Map Seven plot groupings were distinguished by six iterations of TWINSPAN, and named for the one or two species that were most abundant within each group. The Alnu covertype is dominated by green alder shrubs (Alnus crispa), CarS is dominated by tussock sedge (Carex stricta), CaTy is a mixture of sawgrass sedge (Carex lacustris) and cattail (Typha latifolia), Frax is dominated by black ash trees (Fraxinus nigra), SagG is dominated by arrowhead (Sagittaria graminea) and associated aquatics, Scir is a monospecific stand of three-square rush (Scirpus americanus), and Spar is dominated by large-fruited burreed (Sparganium eurycarpum). The Alnu and Scir classes consisted of only one plot each.
The Thiessen polygon map generated from these data appeared somewhat blocky, but revealed the major spatial patterns of vegetation distribution (Fig. 3). Wetlands growing in the bed of the Pokegama River were dominated by the Spar covertype. Spar stands were also common in backwater areas of intermediate water depth. SagG and CaTy occurred as large, contiguous polygons in the deepest and shallowest portions of the backwater, respectively. The levee consisted of eight plots classified as CarS, four plots classified as Frax, and one plot classified as Alnu. The single Scir plot occurred in the riverbed near the northernmost tip of the study site.
The sizes and shapes of each Thiessen polygon were determined by the proximity and location of adjacent sample points. Thiessen polygons define individual areas of influence around each point in such a way that the polygon boundaries are equidistant from neighboring points, and each location within a polygon is closer to its contained point than to any other point (Maggio and Long 1991). Polygon boundaries were close to the sample points where the points were closely spaced, such as between the riverbed and the levee, but far from the sample points in the area behind the convex curve of the levee where there was a gap in the sampling grid (Fig. 3). Each covertype classification is correct at the point where it was measured, but classification certainty decreases with distance from that point. The Thiessen polygons are therefore not necessarily representative of the true spatial extent of vegetation types, particularly in areas where sample points are widely spaced.
CCA/Color Composite Map The first three canonical axes were primarily related to water depth, soil organic matter content, and soil pH, respectively. When the interpolated data layers for each of the three axes were displayed simultaneously as a three-band image, the result was a confusing array of colors that was difficult to interpret (Fig. 4a). Therefore, the data layer for the third canonical axis, which did not substantially increase the cumulative percentage variance explained by the first two canonical axes, was dropped from further analyses.
The first canonical axis separated the levee from the deep backwater, and the second canonical axis separated the low organic matter soils from high organic matter soils. The color composite map for these two CCA axes illustrates these interactions (Fig. 4b). Values for CCA axis 2 were assigned to the red layer for display. Values for CCA axis 1 were assigned to the blue and green layers, which combine to make cyan. Neutrals (whites and grays) on the composite map indicate areas where the values for CCA axes 1 and 2 are similar, white indicating an area where both values are high, and dark gray indicating an area where both values are low. Highly saturated colors (ie. bright red, cyan) indicate areas where the values for the CCA axes are opposite: where one is high, the other is low. Intermediate colors represent intermediate values for each CCA axis.
Colors on the composite map for CCA axes 1 and 2 (Fig. 4b) can be interpreted as follows: light pink, light cyan = no surface water, low to intermediate organic matter soils (ie. levee) dark gray = deep surface water, high organic matter soils bright red = intermediate to deep water, low organic matter soils (ie. riverbed, northern portion of backwater) bright cyan = shallow surface water, high organic matter soils (ie. southern portion of backwater) dark red, dark cyan = intermediate water depth and soil organic matter (ie. central portion of backwater)
The first two CCA axes explained 74.7% of the cumulative variance of species-environment relationships, but only explained 11.6% of the cumulative variance of the plant species data alone, because many of the species were widely distributed across the environmental gradients represented by the CCA axes. The distribution of species such as Carex stricta and Carex lacustris were well explained by CCA axes 1 and 2, respectively, but the distribution of more ubiquitous species (e.g. Sparganium eurycarpum) was poorly explained. Therefore, although this map depicts environmental gradients well, it does not depict species distribution as well. This would not be the case in an ecosystem in which plant communities have discrete species assemblages and environmental requirements.
Airborne Videography Map Nine cover classes were developed from the image analysis and accompanying field work (Table 1). Burreed (BU) and cattail-sedge (CS) were the most abundant covertypes, each constituting 31% of the study site area (Sersland et al. 1995). Other vegetation types of large extent were cattail-burreed (CB) and arrowhead (AC, AT). Only small areas of tussock sedge (TS) and trees & shrubs (TR) were identified (Figure 5).
Table 1. Classes of vegetation distinguished by image analysis, with equivalent TWINSPAN classes.
|CODE DESCRIPTION||DOMINANT PLANT SPECIES||TWINSPAN EQUIVALENT|
|TR||Trees & shrubs||Fraxinus nigra, Alnus crispa||Frax, Alnu|
|TS||Tussock sedge||Carex stricta||CarS|
|AC||Arrowhead, clear water background||Sagittaria graminea||SagG|
|AT||Arrowhead, turbid water background||Sagittaria graminea||SagG|
|CS||Cattail-sedge||Typha latifolia, Carex lacustris||CaTy|
|CB||Cattail-burreed||T. latifolia, S. eurycarpum||Spar|
|SA||Submersed aquatic||Nuphar variegatum, Ceratophyllum||none|
Classification accuracy, determined by comparing the August 1993 field data with the classified multi-date composite image, was about 60% (Sersland et al. 1995). Previous attempts to use conventional color videography for mapping wetland vegetation have been less successful (Jennings et al. 1992). The level of accuracy we achieved by classifying a multi-date composite color image was comparable to the levels of accuracy obtained using more expensive multi-spectral videography (Bartz et al. 1992, Thomasson et al. 1994).
The levee vegetation was particularly poorly identified (Figure 5). Several factors may contribute to this: (1) narrow configuration - in many places, the levee was only as wide as a single tree canopy, (2) leaf-off imagery - the September image was acquired after the leaves had dropped from the trees and shrubs, which would have made them difficult to detect (3) high biodiversity - levee vegetation consisted of a diverse assemblage of species having different spectral reflectance qualities, (4) shadows - the shadows cast by levee trees and shrubs on the July image could have confused the classification. Classification accuracy was improved to 66% correct by excluding reference data points vegetated by trees & shrubs (Sersland et al. in press).
There was good correspondence (52%) between the map generated by image classification and the map generated by Thiessen polygons from TWINSPAN analysis of the field data (Sersland et al., 1995). The two maps corresponded well in the extensive cattail-sedge area in the southern portion of the study site, and in burreed beds scattered throughout the study site. Arrowhead beds also corresponded well in the northern portion of the backwater area.
Both the image map and the Thiessen polygon map have inherent limitations, so one cannot be thought of as more "correct" than the other. The Thiessen polygon map has high classification accuracy at the center point of each polygon, but uncertain classification accuracy elsewhere. The image-derived map has high classification accuracy for certain covertypes, but low classification accuracy for others. The spatial resolution of the airborne video image was much smaller (pixel size = 1 m2) than the spatial resolution of either of the maps generated from field data. As a result, the minimum mapping unit is much smaller, and boundary configurations between vegetation categories appear much more natural (Figure 5).
The use of field data to map vegetation is preferred by most botanists, because they are able to count and measure individual species. Remote sensing is by definition a "hands-off" technique, which is less satisfying to scientists accustomed to studies at the scale of individual species. Remote sensing techniques are less discriminating of background information, such as water turbidity, which could be totally screened out by a field botanist. Given that this background information may be ecologically significant, its detection is not necessarily bad, but it complicates preparation of vegetation maps.
Because floristic data are complex, with many species co-occuring at the same location, analysis and/or interpretation is required in order to classify the data into vegetation assemblages. A traditional approach has been to perform this classification in the field, with the botanist assigning vegetation into pre-determined community types. This approach works well where vegetation occurs naturally in distinct assemblages, but is less suitable where vegetation distributions overlap each other. Statistical methods such as TWINSPAN and CCA provide new tools for distinguishing vegetation assemblages more objectively.
Point data obtained by field measurements are 100% correct at the points where they are sampled, assuming that there is no measurement, classification, or recording error. Constructing a map, however, requires that point data be interpolated. Uncertainty increases with distance from the sample point, so any interpolation technique introduces error. This error is greatest in areas of low sampling density, and in areas where vegetation characteristics change very quickly over space, as in the vicinity of the levee at our study site.
Constructing a map from point data requires collection of locational as well as floristic data, a fact often overlooked by field ecologists. Prior to the development of GPSs, accurate locational data had to be acquired by expensive surveying methods, which required bulky equipment and the presence of a surveyed benchmark in the vicinity. This made conventional surveying very impractical in remote and marshy areas. The development of GPS and the decreasing costs of GPS equipment have made it an indispensible tool for spatial ecology.
Vegetation maps are traditionally categorical, depicting plant groupings as large, discrete entities delimited by distinct boundaries (Fig. 3). Maps derived from remote sensing are also categorical, but the minimum mapping unit is constrained not by sampling density, but by the spatial resolution of the sensor. In the case of our airborne videography, the minimum mapping unit was much smaller and the number of polygons was much larger than that of the more conventional Thiessen polygon map (Figs. 3, 5). Vegetation maps derived from continuous numerical data are more difficult to interpret than categorical maps, but may depict vegetation and environmental gradients more faithfully than do categorical maps (Fig. 4).
In summary, all of the techniques we used produced useful vegetation maps, each with inherent inaccuracies. The technique of choice would depend upon the planned use of a particular map. The TWINSPAN/Thiessen Polygon Map would be preferable for the schematic depiction of discrete vegetation assemblages, the CCA/Color Composite Map would be preferable for depicting vegetation patterns relative to environmental gradients, and the Airborne Videography Map would be preferable for depicting fine spatial detail. Each map is merely a model of the complex reality that exists in nature.
Britton, N.L., and A. Brown. 1970. An Illustrated Flora of the Northern United States and Canada. Dover Publications, New York.
Burrough, P.A. 1986. Principles of Geographical Information Systems for Land Resources Assessment. Oxford University Press.
Fassett, N.C. 1957. A Manual of Aquatic Plants. University of Wisconsin Press, Madison.
Green, P.J., and R. Sibson. 1978. Computing Dirichlet tesselations in the plane. Comput. J. 21:168-173.
Hill, M.O. 1979. TWINSPAN - A FORTRAN Program for Arranging Multi- variate Data in an Ordered Two-Way Table by Classification of the Individuals and Attributes. Cornell University, Ithaca, NY.
Hurn, J. 1989. GPS: A Guide to the Next Utility. Trimble Navigation, Sunnyvale, CA.
Jennings, C. A., P. A. Vohs, and M. R. Dewey. 1992. Classification of a Wetland Area Along the Upper Mississippi River with Aerial Videography. Wetlands, Vol. 12,No. 3, pp. 163-170.
Maggio, R.C., and D.W. Long. 1991. Developing thematic maps from point sampling using Thiessen polygon analysis. pp. 1-10. In Proceedings, GIS/LIS'91, Atlanta, GA, Volume 1. American Society for Photogrammetry and Remote Sensing, Bethesda, MD.
Mueller-Dombois, D., and H. Ellenberg. 1974. Aims and Methods of Vegetation Ecology. John Wiley & Sons, New York.
Palmer, M.W. 1993. Putting things in even better order: the advantages of canonical correspondence analysis. Ecology 74:2215-2230.
Pielou, E.C. 1984. The Interpretation of Ecological Data. John Wiley & Sons, New York.
Sersland, C.A., C.A. Johnston and J. Bonde. 1995. Assessing wetland vegetation with GPS-linked color video image mosaics. pp. 53-62. In: Proceedings, 15th Biennial Workshop on Color Photography and Videography in Resource Assessment. American Society for Photogrammetry and Remote Sensing, Bethesda, MD.
ter Braak, C.J.F. 1991. CANOCO a FORTRAN program for community ordination by [partial][detrended][canonical] correspondence analysis, principal components analysis and redundancy analysis. Version 3.12. ITI-TNO, Wageningen, The Netherlands.
Thomasson, J. A. and C. W. Bennett, B. D. Jackson, and M. P. Mailander. 1994. Differentiating Bottomland Tree Species with Multispectral Videography. Photogrammetric Engineering and Remote Sensing, Vol. 60,No. 1, pp. 55-59.
Carol Ann Sersland Research Assistant Department of Forest Resources University of Minnesota 1530 North Cleveland Avenue St.Paul, MN 55108-6112 Phone: 612-624-3459 Email: email@example.com
John Bonde GIS Manager Natural Resources Research Institute University of Minnesota 5013 Miller Trunk Highway Duluth MN 55811 Phone: 218-720-4266 FAX: 218-720-4219 Email: firstname.lastname@example.org
Deborah Pomroy-Petry Research Plot Technician Natural Resources Research Institute University of Minnesota 5013 Miller Trunk Highway Duluth MN 55811 Phone: 218-720-4357 FAX: 218-720-4219 Email: email@example.com
Paul Meysembourg Principal Laboratory Technician Natural Resources Research Institute University of Minnesota 5013 Miller Trunk Highway Duluth MN 55811 Phone: 218-720-4275 FAX: 218-720-4219 Email: firstname.lastname@example.org