Garrett Schairer Fish and Wildlife Information Exchange Virginia Polytechnic Institute and State University Blacksburg, VA 24060 gschaire@vt.edu

Abstract
 

The Virginia Gap Analysis Program (VAGAP) was faced with the need for more ground truth points to perform our land cover classification and error assessment of remotely sensed Landsat TM imagery.  This need, in conjunction with limited time and money, prompted us to develop a method that was time and money efficient.  Others have labeled land cover in the lab by adding points to an ArcView 3.0a GIS™ shapefile.  We sought to employ this methodology in the field where we could add points without collecting a Global Positioning Systems (GPS) point at each suitable location.  The land cover classes were general enough that we could readily identify them from the road, often without stopping.  The agricultural and forested matrix found in the piedmont and coastal plain of eastern Virginia made data collection fast because we could see beyond many fields into forested types and could add points in each of the classes.  In an area with a similar land use/land cover composition this method could be employed to rapidly collect ground points.

We were able to collect over 300 points in an 8-hour day.  Compared to other methods, it was determined that this method was extremely cost and time effective and was possible due to our relatively general Anderson Level I classification scheme (Anderson 1976).  This technique reduced data processing time in the lab while eliminating the "paper trail" typically associated with ground truth data collection.  At this level of categorical specificity, the increased efficiency and timely transfer of data made it possible to collect enough points to perform our remote sensing classification and error assessment in eastern Virginia.

Introduction

To table of contents

Ground truth data collection is a vital part of any remote sensing effort.  The Virginia Gap Analysis Program (VAGAP) has been using a combination of aerial photographs, digital maps, and aerial videography to perform classifications and error assessments of Landsat TM imagery.  However, we determined that the number of known ground points available to us was limited and we would need more points to perform this basic level of classification and subsequent error assessment.  Consequently, we set out to collect more points within the study area.  Working within the constraints of limited time and money, we devised a methodology that maximized the number of collected points without expending all our resources.  We feel we have a methodology in which we can rapidly collect hundreds of points using technology that not only makes these points spatially explicit but also removes some of the problems associated with other data collection efforts.  
 

Methods

To table of contents

For this classification we used a rather general classification scheme (Table 1) similar to an Anderson Level I classification (Anderson et al. 1976).  In January and February 1998, we performed "windshield surveys" in which we drove portions of the state collecting points of known ground cover in 6 cover types.  We had two individuals in the car, one drove, and one recorded known points.  We used a 75 MHz Pentium laptop running ArcView 3.0a™ for data collection.  ArcView projects were created that contained roads and water themes as well as our raw Landsat TM imagery.  A point theme was created to which points of known ground cover were added, and an identifying land cover attribute was added in the associated theme table.  As we drove, one observer followed the vehicle's position on the digital roads theme while also using the water theme and the raw Landsat imagery to aid in finding our position.  The roads theme available is very extensive, making it extremely easy to follow the vehicle's position on the laptop.  When one of the desired classes of land cover was encountered, a new point was added to the points theme.  In the table associated with this theme, the class code would be added, and the observer would return to scanning the landscape for the next point.  If at any time the observer "got lost," we stopped the car and collected a GPS location.  Though this GPS location was uncorrected, we could use the rough position in conjunction with the features on the various themes to relocate our position.

By the end of each field session, we had created one large collection of points that could be manipulated in ArcView™ in the lab.  The coordinates of the points could be added in the lab using an ArcView script that is available (addxycoor.ave).  Since our TM imagery was in Universal Transverse Mercator projection (UTM), these points were in the correct projection and required little further pre-processing.  With the data in digital form, all traditional data manipulation tools are available as well.

Results

To table of contents

Over a 6-day sampling period, 1505 points were collected.  These sites were remote, requiring some travel time to get to the study areas and reducing our collection rates.  Still, we averaged 251 points per 8-hour day and collected in excess of 300 points in 8 hours on some days.  Certainly some training is required to get to a collection rate as high as this; however, we feel that gaining some familiarity with the collection methods is necessary for any efficient data collection effort.  Our aerial videography project has been able to label 80-100 points per day in the same physiographic region after approximately 100 hours of training (VAGAP unpub. data).  The detail level of our aerial videography is greater, requiring some more time to accurately record the land cover.  Comparatively, we were able to collect more points per hour or dollar in our select classes than the videography effort could collect.

 Figure 1.  An example of a path of travel.  The points in purple are those that have been added to the theme.  On the left, the table is visible, where the attribute for that point is added.  The point with an arrow indicates one of those times when we were able to label a pasture class (A) and the 
deciduous forest (B) behind it without leaving the road.

 
 Discussion

To table of contents

We feel this is a superior data collection method for some applications.  For the classification effort we were performing, this method was very fast and made collecting hundreds of points much easier.  Rather than having to stop and collect a GPS point at each location, we were able to collect points while moving, which made data collection faster and easier.  Additionally, by eliminating the need for being directly "on the GPS spot," we were able to add points that were up to 300 meters off the road without having to go out into that cover type.  This proved extremely useful within the study area because most of the region consisted of a matrix of private agricultural and forested lands in eastern Virginia.  For example in Figure 2, we were able to put a point in a pasture from the road and then label a point in the deciduous woods behind the pasture without leaving the car!

Also, by eliminating the GPS point associated with each location, we were able to collect on any day without concern for base station blackout days.  No effort was required to get GPS base station data and differentially correct the GPS points.  This made these points immediately available for classification and accuracy assessment.  These points could be split by area (i.e.: for different scenes) or class (for stratification purposes), or simply exported to a database file that is importable into other remote sensing packages or accuracy assessment programs.  We used PCI EasiPace™, TNT Map and Image Processing Systems™ (MIPS) software, and a separate accuracy assessment program (ACC4WIND™).  The ability to export these points from ArcView™ as a dBase™ file made the transition between software packages extremely fast.
 

 Figure 2.  An example of where we were able to add two points to our point theme without leaving the road surface.  The near point was added as a pasture/hay class, while the further class was added as a deciduous forest.

Another advantage of this method was having Landsat TM imagery on the laptop, which enabled the observers to determine the effective size of the patch of land cover.  We could see whether a patch was visible on the imagery, eliminating the guesswork associated with patch size determination while resolving some mixed pixel/edge problems.

For this level of categorical specificity in land cover, these methods appear superior to aerial videography.  Our early air video data were error prone due to the potential tilt, yaw, and roll of the plane at the time of collection.  Later analysis showed the location of the labeled points, which were tied to corrected GPS point, were off by an average of 70 meters (Virginia Gap Analysis Aerial Videography Page, 1998).  Due to this error, our early video points were only usable in larger homogeneous areas, eliminating many known points from consideration.  Our methods resolved some of these issues that restricted our early videography efforts.

Finally, simply entering attribute data while in the field eliminates time spent entering and checking the data at a later time.  In all, it is an extremely effective method of data collection for remote sensing projects with a general level of categorical specificity.
 

Reference

To table of contents

Anderson, J. R., E. E. Hardy, J. T. Roach, and R. E. Witmer. 1976. A Land Use and Land Cover Classification System for use with Remote Sensor Data. U.S. Geological Service Professional Paper 964. 28pp.

Virginia Gap Analysis Program Aerial Videography Page. ed. S. McNulty.   Jan 6, 1998. Fish and Wildlife Information Exchange, Virginia Polytechnic Institute and State University. <http://fwie.fw.vt.edu/WWW/video/>