GPS REFERENCED AERIAL VIDEO FOR ACCURACY ASSESSMENT OF A SMALL-SCALE LAND-COVER MAP

1. Introduction







Studies of global climate have generated the need for small scale



land-cover maps.  Analysis of Advanced Very High Resolution 



Radiometer (AVHRR) imagery has been proposed as a preferred



method for creating these maps (Townshend et al., 1991).  



Accuracy assessments of AVHRR derived map products have been



limited to comparisons against existing maps of unknown accuracy



or against a limited number of field observations (Loveland et



al., 1995).  Traditional accuracy assessment techniques rely on



comparing a sample of ground observations against the map



product.  These techniques are sufficient when working on large



and medium scale data sets, such as those derived from aerial



photography or Landsat imagery, but become problematic when



applied to small scale data sets, such as those derived from



AVHRR imagery.   In most cases, field observations are limited to



point samples, which may represent only a very small area within



a single resolution element.  Terrain conditions often constrain



ground observations to areas accessible by roads and trails. 



These sites may not be representative of land-cover over larger



areas.  Extensive travel is required to acquire a large enough



sample to be representative of the large areas needed for global



studies. As a result of these limitations, field surveys are



difficult to justify since global scale land-cover classes are



very general and therefore hard to extrapolate to the large



mapping units.







The objective of this study was to develop a method for rapidly



assessing a generalized land-cover map by integrating aerial



video and  the Global Positioning System (GPS).   Light aircraft



provide a means for covering large distances in a short length of



time.  Terrain does not limit access except in very mountainous



areas and road access is never an issue.  Observations made from



aircraft offer a means for collecting land-cover information in



areas which could not otherwise be visited (Marsh, et.al. 1993). 



Video provides a technical means for recording and replaying



imagery collected from light aircraft.  With the incorporation of



position information from a GPS receiver, it is possible to use



aerial video to collect efficient transect samples of land-cover



over long distances and rugged terrain. 







An aerial video accuracy assessment was performed for a land-



cover map of Arizona.   A rule-based classification model,



implemented in AML, was applied to three years of AVHRR imagery



(Kliman 1996).   The AVHRR imagery was derived from the EROS Data



Center (EDC)  Conterminous U.S. AVHRR Biweekly Composites on CD-



ROM, projected in Lambert's Azimuthal Equal Area (LAZEA)



projection, with 1 km grid cells.  The output of the model is a



single ArcInfo GRID file summarizing the results.  This GRID



file serves as the basis for a small scale map, entitled the



Summary Map (figure 1).  







The land-cover map was derived through a



rule-base classification of  biweekly NDVI and thermal imagery



spanning 1990 to 1992.   Land-cover was classified according to



the scheme established by Brown and Lowe for their 1973  Natural



Vegetative Communities of Arizona map (figure 2).  







An accuracy assessment was also made for the Brown and Lowe map in order to



provide a comparison between the two maps and the video observations.







2. Field Methods







Two flights were made, totalling 2171 km in length (figure 3). 



The flights were planned to overfly as many different land-cover



types as possible in the minimum time.  







The first was a short, exploratory flight over southeastern 



Arizona on April 18, 1995.  The second flight took place on August 



1 and 2, 1995.  This flight spanned the state, with a round-robin 



route from Tucson to Holbrook to Page to Bullhead City returning 



to Tucson.  A Cessna182, piloted by the author, served as the platform.  



The video camera was a Cohu 1310 series color video camera, with 



a 12.5 mm lens.   GPS coordinates from a Trimble Pathfinder Basic Plus 



was stamped onto each video frame using a Horita GPT-50 unit.  The



video was recorded in 8mm format.  The flight path was recorded



by the Pathfinder. 







Flying height was established at a nominal altitude of 460 m



above ground level (agl).  An exception to this was the portion



of the flight through the Grand Canyon Special Flight Rules Area



(SFRA), where regulations require a minimum altitude of  3200 m



above mean sea level (msl).   Altitude above ground level was



estimated visually and verified by comparing barometric altitude



against ground elevations depicted on 1:500,000 aeronautical



charts. 







The Cohu camera has a field of view of 30o.  For these flights,



the camera was oriented at a low oblique orientation,



approximately 15o off nadir.    In this configuration, the camera



has a spatial nominal resolution of  0.70 m at the bottom of the



frame and 0.81 m at the top of the frame.  The swath width is



approximately 265 m.  While this lacks detail necessary to



identify plant species, it is adequate for discerning the



relative sizes of shrubs and  trees, and  qualitatively



estimating plant densities.  







Positions were not differentially corrected, giving a nominal



spatial resolution of 100 m.  The GPS receiver records the



position of its antenna, which was affixed to the rear window of



the aircraft.   At  460 m agl with 15o oblique view, the center



of each video frame is offset 119 m to the left of recorded



position of  the flight path.  For the purposes of comparison



with a classification derived from 1 km AVHRR data, this spatial



resolution is more than adequate.  Each video frame is titled



with the GPS position (figure 4).  Latitude and longitude are



depicted in degrees and decimal minutes.  Altitude is recorded in



feet above mean sea level. 









3 Analysis







After the flying was completed, the tapes were analyzed in the



laboratory.   As the tape played, the positions of transitions



between land-cover types were recorded.  Land-cover type was



visually  interpreted from the video. Where the transition



occurred over a distance of several kilometers, a video frame



representing the center of the transitional segment was selected



to assign the position.   The video equipment did not operate



perfectly during the survey.  Technical problems resulted in gaps



in the video coverage, resulting in gaps along the flight lines.  



Out of a total of 2171 km flown, there were 1539 km of usable



video data.  This coverage represents a sample of  354 km2, 0.12%



of the total area mapped.  All land-cover types, with exception



of  Tundra and Mt. Meadow, were identified in the video data.







Identifying vegetation type to the community level is an



uncertain process.  Many of the classes defined by Brown and Lowe



are separated by the presence or absence of specific plant



species.  With few exceptions, positive identification of a plant



species is based on the recognition of specific physiological



traits, such as the shape of leaves or seeds.  These features are



not visible in the aerial video.   The video interpretation had



to be based on the recognition of gross features, such as trees,



shrubs or open grassy spaces, in association with a-priori



knowledge about the physiographic and geographic distribution of



land-cover within the state.   In this case, the interpreter had



the benefit of making first-hand observations during the video



survey and familiarity with the Brown and Lowe map. 







Some classes were easy to identify.  The forest classes, with



distinctively shaped spruce and pine trees were easy to



recognize.  The deserts and woodlands, in contrast, required a-



priori knowledge to classify.   The Mohave, Great Basin, Sonoran



and Chihuahuan Deserts were nearly indistinguishable from one



another in the video imagery.  J-P Woodland, Chaparral and Oak-



Pine Woodlands were similarly difficult to distinguish.   A



knowledge of their geographic and physiographic limits were



necessary to assign a class based on the location along the video



transect.  It was possible to identify the deserts with some



confidence because they have large homogenous areas with little



spatial overlap.   In contrast, the woodland classes could not be



identified as confidently.  The three classes overlap in both



geographic extent and elevation, and occupy relative small areas



between the deserts and forests.







The video interpretations were entered into the GIS for



comparison with the Summary Map and the Brown  and Lowe.   The



point locations of the land-cover transitions were plotted on top



of the flight lines.   The flight lines were broken into segments



at the transitions.  Each line segment was assigned an attribute



for land-cover type.  The video transects were subsequently



overlaid with the Summary Map and the Brown and Lowe map,



breaking the transect into a larger number of segments, each



carrying an attribute for video interpretation, Brown and Lowe



and Summary Map land-cover class (figure 5).







4. Results







A comparison of segment lengths (table 1) shows the difference



between the occurrence of classes along the transects.   Transect



lengths were more similar between the video classification and



the Summary Map than between the video and the Brown and Lowe



map.  The most conspicuous disagreement is present in the lengths



of the Grassland class.  The majority of this  disagreement



represents the portion  of the transect between Holbrook and



Page.  Brown and Lowe classify this area as Grassland.  The video



observations do not support his classification; there are



numerous shrubs visible in the video throughout this areas, as



well as large areas of "badlands" - eroded and dissected surfaces



covered with low shrubs.





Table 1.  Length of video transect by land-cover class





Land-cover        Video (m)   Summary Map(m)   Brown&Lowe(m)



Fir Forest         21445            28588           22940



Conifer            75581           116657           55274



J-P Woodland      102516            80133          175317



Chaparral          36217            61579           45032



Oak-Pine Woodland  46018            81511           46429



Grassland         156170           128853          541891



Mt. Meadow             0             2046               0



Great Basin       447479           400150          228886



Mohave            309051           290336          223418



U. Sonoran        117938           121545          146108



L. Sonoran        130219           123833           36209



Chihuahuan         96694           104099           17825



Total            1539328          1539328         1539328







A second source of disagreement is the presence of class 8,



Mountain Meadow in the Summary Map classification of the



transect.  This class is not obviously present in the video



imagery nor in the Brown and Lowe map along the transect.  There



is, however,  small area of this class in the Summary Map which



is overlaid by the video transect.  This area is located in the



center of the Kaibab Plateau.  The Brown and Lowe map depicts a



narrow strip of Grassland running down the middle of the plateau. 



The corresponding segment of  video is classified as Fir Forest. 







Confusion matrices were used to quantify the relationship between



the video interpretation, Summary Map and the Brown and Lowe



classes found along the video transect.  Values for percent



agreement and Cohen's K coefficient of agreement were calculated



(Rosenfield, G.H. and Fitzpatrick-Lins, K. 1986).   The



comparison of the video interpretation with the Summary Map



showed an agreement of  65%  (K = .5944) (table 2).  The greatest



source of confusion was between classes 4, 5 and 6, which are



woodland classes.  This result was not surprising, given the



difficulty in interpreting these classes from the video.  The Fir



Forest and Conifer classes had the best agreement.   The desert



classes also showed good agreement.





                Table 2. Agreement between video and maps





Comparison                        Percent      K



Video - Summary Map 13 Classes    65.49%   0.5944



Video - Brown and Lowe            52.92%   0.4598



Summary Map - Brown and Lowe      42.42%   0.3476





The video showed better agreement with the Summary Map than it



did with the Brown and Lowe map. The confusion matrix and



analysis of errors of commission and omission show an overall



agreement of 52%  (K =.4598), less than those observed between



the video and the Summary Map.   Much of the disagreement results



from the area of Grassland identified by Brown and Lowe in the



northeastern corner of the state.  As above, the woodland classes



were also a source of disagreement.  







Agreement between the Summary Map and Brown and Lowe classes



along the transects was also tested.  The results showed an



overall agreement of only 42%  (K of .3476). Disagreement between



woodland classes and between desert and grassland classes was the



source of the low values.   The values are lower than the overall



agreement measures for the entire area of the two maps,



indicating that the transect passed through a higher than



representative number of disputed areas.  Part of this higher



level disagreement may have been the comparatively large segment



of the transect crossing  over the northeastern part of the



state.  Proportionally less data was collected over the



southwestern part of the state, where the two maps showed better



agreement.







The  best agreement is between the video transects and the



Summary Map for all 13 classes.   These results are an



encouraging indicator of the relative accuracy of the rule-based



model in separating land-cover classes.  Unlike the Brown and



Lowe map, which served as the spatial basis for the rule-based



model, the video interpretation provides an independent source of



land-cover information.   While  65.54% is a low level of



agreement relative to conventional means of land-cover mapping



and verification, it provides an indication that there is



positive agreement between video and AVHRR land-cover



classification. 







5. Comparison with field observations







In addition to the aerial video flights, a ground survey was



made.  A total of 19 ground observations were made, representing



nine of the 13 Brown and Lowe classes (figure 6).  







Sites were located throughout the state, I-8 from Tucson to Yuma, 



I-10 between Phoenix and Blythe, I-17 between Phoenix and Flagstaff



and state highways between Tucson and Springerville. 



Observations were made near the two highest points in the state:



one at Apache Peak in the White Mountains, and one near



Humphery's Peak in the San Francisco Mountains; however, it was



not possible to get into the Tundra zone due to snow cover.  The



range of feasible ground visits did not extend far enough



northeast or northwest to include Great Basin or Mohave, nor far



enough southeast to record Chihuahuan Desertscrub.  







Ground sites were chosen on the basis of safety, accessibility



and identifiability.  Many of the sites were along highways and



roads.   Each site was photographed with a 35 mm camera. 



Geographic coordinates were recorded using a Garmin hand-held GPS



receiver.  Two units were used, a model 55 and a model 40.  Both



are autonomous units with a spatial accuracy of approximately 100



m.







The ground observations were entered into the GIS, where they 



were compared against the Brown and Lowe and the Summary Map



(table 3). There were nine sites that showed disagreement with



either the Summary Map or the Brown and Lowe map.  Disagreement



may be the result of misclassification of the field site,



misclassification by the model, or misclassification in the Brown



and Lowe map.  





              Table 3. Agreement between field sites and maps     







Comparison                        Percent   Number



Field -  Summary Map               68.42%     13



Field - Brown and Lowe             68.42%     13



Summary map - Brown and Lowe       63.15%     12



(Note: The low number of sample sites sample prohibited the   



calculation of  K.)







The results showed that  68%, of the sites agreed between the



field observations and Summary Map.  This agreement value is



within 3% of the agreement value of 65% determined from the



video.  The same proportion of field sites, 68%,  agreed with the



Brown and Lowe map, although the 13 sites in agreement were not



the same as those which agreed with the Summary Map.  This



agreement value is higher than the 52% agreement value between



the video transects and the Brown & Lowe map.  Agreement between



the Brown and Lowe and Summary Map classification was limited to



12 of  19, 63%, with the field observations.   This value is over



20% higher than the 42% agreement between the comparison of the



two maps along the video transect.  







The field sites yielded agreement values which were similar for



the comparison between the video transects and the Summary Map,



but were lower for the comparisons with the Brown and Lowe map. 



The similarity of agreement between the video transects, the



field observations and the satellite derived land-cover map



validates the concept that video transects can be used in place



of field observations, with similar results.  The lower agreement



between the video, the Brown and Lowe map, and between the



Summary Map and the Brown and Lowe map is attributed to the lack



of field observations in the northeastern portion of the state,



where the large contiguous discrepancy between the Grassland in



the Brown and Lowe map and Great Basin Desertscrub in the Summary



Map is located.







Conclusions







Aerial video provides a means for overcoming some of the



difficulties inherent in field accuracy assessments.   A large



transect sample can be acquired over long distances in a short



time.    A drawback is the lack of  detail present in the video



imagery.  Land-cover interpretation from the video imagery is



less certain than field observations.  Only generalized



observations of plant size and density can be made.  



Interpretation of  specific land-cover classes requires a-priori



knowledge of the geographic and physiographic characteristics of



the land-cover classes.  







Despite the experimental nature of video land-cover assessment,



this application demonstrated that its utility for validating the



results of  rule-based classification of AVHRR imagery.  The



Summary Map is the first example of a regional scale AVHRR land-



cover classification to be assessed using video.    The



comparison between the video interpretation,  the Summary Map and



Brown and Lowe showed that the best agreement was between the



video and the Summary Map.  Similar levels of agreement were



observed between the field observations and the Summary Map. 



These similar levels of agreement indicate that  video assessment



offers a viable tool for assessing global scale land-cover maps



and may be used in lieu of traditional field based methods. 







Both techniques of rule-based classification and video assessment



are in early stages of development.  The level of agreement



suggests that additional refinement of each technique can yield



even better results.   The assessment for this study was done



with very simple equipment.   More work is needed in both the



technical and procedural areas.  Improvements in image quality



are needed to increase confidence in interpretation.   These



improvements might be made by using a larger format video camera,



a lens with zoom capability, and by incorporating bi-spectral and



multi-spectral video sensors.   GPS position can be made more



precise with differential corrections. More representative



sampling techniques using GIS for flight planning need to be



developed as well.   As advancements in video and GPS technology



make these refinements possible, GPS referenced aerial video will



likely become the technique of choice for assessing the accuracy



of small scale maps.







References



Brown, D.E. and Lowe, C.H. 1973. The Natural Vegetative



Communities of Arizona. 1:500,000 Map. Phoenix: Arizona Land



Resources Information System.





Kliman, D.H. 1996. Rule-based Classification of Hyper-temporal,



Multi-spectral Satellite Imagery for Land-cover Mapping and



Monitoring.  Ph.D. dissertation University of Arizona: Tucson.





Loveland, T.R., Merchant, J.W., Brown, J.F., Ohlen, D.O., Reed,



B.C., Olsen, P. and Hutchinson, J. 1995. Seasonal Land-Cover



Regions of the United States. Map Supplement.  Annals of the



Association of American Geographers. 85(2): 339-355.





Marsh, S.E., Walsh, J.L., and Sobrevila, C. 1993.  Evaluation of



Airborne Video Data for Land-Cover Classification Accuracy



Assessment in an Isolated Brazilian Forest.  Remote Sensing of



Environment.  48:1-25.





Rosenfield, G.H. and Fitzpatrick-Lins, K. 1986. A Coefficient of



Agreement as a Measure of Thematic Map Classification Accuracy. 



Photogrammetric Engineering and Remote Sensing.  52(2):223-227.





Townshend, J.R.G., Justice, C.O., Li, W., Gurney, C., and



McManus, J. 1991. Global Land Cover Classification by Remote



Sensing: Present Capabilities and Future Possibilities. Remote



Sensing of Environment. 35:243-255.



    



First Name: Douglas



Last Name: Kliman



Title: Research Associate



Organization: University of Arizona Office of Arid Lands Studies



Mailing Address: 1955 E. 6th Street Tucson, AZ 85719



(520) 621-8064 office



(520) 621-3816 fax



dkliman@arsc.arid.arizona.edu