Point Sampling Surveys with GPS-logged Aerial Videography

Obtaining sufficient geographically unbiased data for verification and validation of vegetation communities is one of the greatest challenges in developing vegetation base maps for the Gap Analysis Program. These independent data are essential for classifying the Landsat Thematic Mapper (TM) imagery used in all Gap Analysis projects and for assessing the accuracy of the vegetation maps. Low altitude aerial surveys, combined with video data systems that tag each video frame with geographic coordinates from a global positioning system (GPS), provide a cost- and time-effective method for obtaining high resolution data on vegetation communities over large geographic areas. The Gap Analysis Program in New England is using this technology in conjunction with the hyperclustered, multitemporal Landsat TM imagery distributed through the Multi-Resolution Land Characteristics Consortium (MRLC) to produce its vegetation map of southern New England (Slaymaker et al. In press).

Aerial point sampling was developed to characterize the land cover of a region by interpreting a distributed sample of large-scale aerial images (Norton-Griffiths et al. 1982, Dunford et al. 1983). The Arizona Gap Analysis Program used this approach first to interpret its statewide Landsat TM coverage (Graham 1993), using aerial videography in combination with GPS-logged time code. The Arizona project used a Super-VHS camera flown at 2,000 ft above ground that covered 0.5 km at wide angle and zoomed to 15x magnification once every 10 seconds to collect a point sample image 30 m wide. The New England Gap Analysis Project modified Arizona's pioneering system by using two Hi8 band video cameras attached to a portable mount that can be operated from any highwing Cessna. The mount is clamped to the open window frame, then cranked out and adjusted to vertical with a bubble level. The video cameras are mounted vertically beside each other. One is set at wide angle, the other at 12x zoom, providing a swath of 30 m wide large-scale imagery down the middle of 0.4 km wide-angle coverage when flying 600 m above ground level. This approach provides more flexibility than a single camera in both the selection and density of sample points. Geographic position data are recorded in-flight from a GPS unit to a laptop computer using Geolink software. Flight lines can also be entered into this system for navigational purposes and will appear in correct relationship to the plane's position on the computer screen during filming. Time code is "sipped" from the GPS data stream by a Horita GPS time code generator to provide a matching SMPTE time code for the video tape recorders. SMPTE time code is the standard audible timing signal recorded to the audio track of professional video. Horita's time code generator substitutes GPS code for the normal internally generated signal, allowing each frame to be matched to a geographic position. In our system, time code is recorded directly onto the video images as well as the audio track, simplifying the synchronization of the two tapes during playback. We flew 10 - 24 km spaced transects of all six New England states (3,000 km) in the spring and fall of 1994, so as to capture both phenologies of our deciduous forests.

The video tapes from these transects are used to select and label sample Landsat pixel data. Two TV monitors, one each for the wide angle and zoom videos, are set up beside a computer monitor showing the corresponding portion of the Landsat image. The GPS flight data are overlaid on the Landsat image as vector points that can be queried for their time code, allowing the video frames to be matched to that image. As the video tape is interpreted and plant communities identified, specific pixels in the Landsat image are tagged with their forest type or vegetation class. These points are later extracted as a set of attributed coordinates and used as training sites in a supervised classification or, as in our case, systematically modeled for a set of inference rules to relabel the hyperclustered classification. Each selected pixel takes only seconds to tag, and we collect 18,000 or more points per image in a stratified sample by region and by topographic slope. One quarter of the sample points (stratified by vegetation type) are set aside and later used to access the accuracy of the final vegetation map. The remaining points represent only 0.06% of the total pixel population of a Landsat image, but this sample is sufficient for modeling of the probable vegetation types of each spectral class under different conditions of terrain and spectral mixtures.

The models for each vegetation type are developed with a set of Excel templates. The contingency tables sort the sample points by slope/aspect, frequency of appearance within a vegetation type, and the characteristics of their immediate neighborhood (25 pixel block) within those vegetation types. These data are then used in another set of Excel templates to construct inference rules that relabel each Landsat pixel to its most probable vegetation class for its location and spatial context. The templates are available on our World Wide Web (WWW) home page (see "New England GAP Analysis" at http://tove.fnr.umass.edu/gaphome.html, or ftp://tove.fnr.umass.edu/pub/gap), along with complete sets of our rules, a more detailed explanation of the process, and our initial accuracy assessments, which indicate a near 90% reliability for the seven forest types tested so far.

Technical assistance, including on-site workshops, acquisition of aerial video coverage, and assistance in setting up video interpretation stations, has been provided to a variety of Gap Analysis projects such as Colorado, Florida, Maine, Montana, Oregon, Tennessee, Vermont, West Virginia, and parts of California, as well as new applications of GAP methods in Madagascar, Mexico, and Portugal. Several GAP state projects now have aerial video camera systems and are using them cooperatively with other states. Contact Dana Slaymaker at the University of Massachusetts at (413) 545-4853 or dana@tove.fnr.umass.edu for additional information or technical assistance on aerial videography and interpretation of multiseasonal hyperclustered TM data.

Literature Cited

Dunford, C., D. Mouat, M. Norton-Griffiths, and D.M. Slaymaker. 1983. Remote sensing for rural development planning in Africa. The Journal for the International Institute for Aerial Survey and Earth Sciences 2:99-108.

Graham, L.A. 1993. Airborne video for near-real-time vegetation mapping. Journal of Forestry 8:28-32.

Norton-Griffiths, M., T. Hart, and M. Parton. 1982. Sample surveys from light aircraft combining visual observations and very large scale color photography. University of Arizona Remote Sensing Newsletter 82-2:1-4.

Slaymaker, D.M., K.M.L. Jones, C.R. Griffin, and J.T. Finn. In press. Mapping deciduous forests in southern New England using aerial videography and hyperclustered multitemporal Landsat TM imagery.

Dana M. Slaymaker
New England Gap Analysis Project
University of Massachusetts, Amherst