Interpretation of Aerial Videography in Western Virginia Using ArcView

---

• Introduction
• Methods
• Field Verification and Interpreter Training
• Interpretation
• Results
• Discussion
• Summary
• References

INTRODUCTION

The western region of Virginia includes parts of three physiographic provinces: from east to west, they are the Blue Ridge, Ridge and Valley, and Appalachian Plateau (Woodward and Hoffmann 1991). The vegetative communities in this region historically consisted of forests whose composition depended upon slope, aspect, and elevation. After four centuries of human modification, the current landscape is a complex mosaic of agricultural lands and forest patches.

Efforts to map the land cover in western Virginia using satellite image classification have been confounded by topographic variation and species diversity. An extensive dataset of known ground cover is needed to identify land cover classes, but collection requires a large allocation of time and money. The Virginia GAP analysis project (VA-GAP) chose aerial videography to quickly collect data. The cost-effectiveness and efficiency of aerial video is well-known (Graham 1993, Slaymaker et al. 1996). Aerial video data is being used for generation as well as accuracy assessment of the land use/land cover map, based on Landsat Thematic Mapper (TM) and ancillary data (McCombs et al. 1997) The goal was to collect 100 data points per Supertype for each TM scene.

We investigated the feasibility of interpreting aerial video to VA-GAP's Supertype level. Supertypes are based on assemblages of Society of American Foresters (SAF) types (Table 1, for a complete version see the Virginia GAP Vegetation Classification ) and were chosen because some herbaceous and understory species could not be identified to the Alliance level with aerial video. In this poster, we illustrate the specific procedures used in interpreting western Virginia with ArcView 3.0(tm) software. Finally, we address the utility of interpretation using the ArcView aerial video system.
 

• I.A. EVERGREEN FOREST (>60% Conifer) [12]
• I.A.8.N. Temperate And Sub-Polar Needle-Leaved Forest (Natural)
• I.A.8.N.b. Evergreen, needle-leaved forest with rounded crowns
• I.A.8.N.b.i. White Pine Forest
• I.A.8.N.b.i.21 Eastern White Pine 108
• I.A.8.N.b.i.22 White Pine, Hemlock 109
• I.A.8.N.b.ii. Montane Yellow Pine Forest 106
• I.A.8.N.b.ii.79 Virginia Pine 105
• I.A.8.N.b.ii.01 Table Mountain Pine 106
• I.A.8.N.b.ii.45 Pitch Pine 107
• I.A.8.N.b.iii. Submontane Yellow Pine Forest 103
• I.A.8.N.b.iii.80 Loblolly Pine, Shortleaf Pine 103
• I.A.8.N.b.iii.79 Virginia Pine 111
• I.A.8.N.b.iii.81 Loblolly Pine 112
• I.A.8.N.b.iii.70 Longleaf Pine 113
• I.A.8.N.b.iii.75 Shortleaf Pine 114
• I.A.8.N.b.iii.98 Pond Pine 115

To table of contents

We followed a modified procedure based on Slaymaker et al. (1996). Video was recorded from a small fixed-wing aircraft with two cameras. One camera was set for a wide angle view, approximately 0.5 km wide, and the other was zoomed to show a 30m wide swath at the center of the wide angle view (Slaymaker et al. 1996). Two transects were flown during Fall 1996, one over the southwest and a second in the west-central mountain region (Fig. 1). The combined length of the transects was 1330 km. Click on the smaller images below to view the full-size figure.
 

Figure 1. Western flightlines. Figure 2. Video frame. Figure 3. ArcView setup. 
 

We differentially corrected and exported the point file to ArcView. This file was overlaid onto five leaf-off 1992 and 1993 TM scenes. Figure 3 shows the ArcView display with the TM scene in the background and pink GPS points overlaid onto the TM scene. The optimal combination of TM scene spectral bands was 3, 4, and 5 displayed in the Blue, Red, and Green channels, respectively. Because ArcView does not have sophisticated multi-band image enhancement, it was necessary to manually adjust the image's bands to visually improve the difference between land cover types. Ancillary datasets such as roads and rivers were also displayed.

Field Verification and Interpreter Training

Interpreters selected sites for field verification of land cover types. Sites were located near roads to ensure access. We used Snappy(tm) software to "grab" video images on a monitor and convert them to digital images. At each site, we identified land cover, aspect, slope, and presence of streams or water. If the site was forested, we labeled individual trees in the image and listed the dominant overstory and understory species and closest cover type.

Field-verified images were incorporated into a video key. All interpreters trained with the key to minimize differences in interpreter ability from prior knowledge of native trees or aerial photo interpretation. Once trained, interpreters could discriminate between most tree species using crown color, shape, and texture. This key was used during interpretation as a quick on-screen guide.

Interpretation

Land cover was interpreted using the VA-GAP Land Use/Land Cover Classification (Table 1). We synchronized the time code on the video frames to the corresponding geographic location in the point database. The video was played until the interpreter encountered a homogeneous land cover type (e.g., oak-pine forest). We restricted interpretation to parcels >1 ha to limit spatial error due to the rolling motion of the plane and camera (Graham 1993).

We noted that error also could be introduced by spectral mixing of different cover types (forest-field edges) and misalignment of the satellite image with the ancillary datasets. To address this, we interpreted a land parcel by adding a point close to the center of the parcel rather than labeling the point itself. We labeled the dominant tree species, land cover code, and descriptive site characteristics in the attribute table (Fig. 3). Land cover types were interpreted to the Alliance level if possible and were later aggregated to the Supertype level and catalogued by TM scene.

RESULTS

To table of contents

We visited 47 field sites and compiled 268 images. Training two interpreters required 300 hours, and field verification required 104 hours. We calculated the deviation of the video location from the GPS point in the transect. Average deviation due to airplane pitch and yaw was 70 m with a range of 0-8 pixels (0-240 m).

We interpreted 1357 points in the western mountains of Virginia (Table 2). Approximately 32 points were interpreted per day. Forty nine percent (49%) of the points were labeled deciduous land cover types, 8% were coniferous, 6% were mixed, and 29% were agricultural, with the remaining 8% comprised of shrub, wetland, water, and disturbed types. We obtained over 100 points for two TM scenes in the Red/White/Black Oak and the Pasture/Field Crop Supertype, while the remaining classes had fewer than 100 points each.

DISCUSSION

To table of contents

We successfully generated land cover data for western Virginia using aerial video. We now have detailed transect data on tree species distribution for previously unsampled regions and areas inaccessible on the ground. At the Alliance level, the number of points was not sufficient for classification of the satellite images. We interpreted few points in the mixed and shrub classes, most likely due to the difficulty in finding parcels larger than 1 ha of these naturally transitional cover types. We encountered few disturbed areas or wetlands on our transects.

Interpretation was faster and cheaper than our field verification. However, our interpretation rate was lower than expected. The initial location of the site was crucial, because of error introduced from the plane motion during flight (Slaymaker et al. 1996). Interpretation of the video required 3-8 minutes, and identification of tree species within diverse deciduous forests took the most time.

We collected fewer points than anticipated because of two factors. The mixture of almost 70 land cover types limited interpretation of heterogeneous areas. Restricting interpretation to 1-ha parcels limited data collection in some areas. Despite this, we feel the 1-ha parcel limitation is robust and will allow us to exclude variation from edges of different cover types in the satellite image classification. We will continue to acquire interpreted land cover data for these TM scenes using additional video to reach the goal of 100 points per class per scene.

The interpretation system we created in ArcView functioned well and was easy to learn. Another advantage is the ease of sending products to other users or exporting data to other formats with minimal alteration. The main disadvantage of using ArcView was the lack of fine image controls for adjusting the TM scene that are available with other image processing packages. Once set up, our system efficiently integrated image display and database management.

SUMMARY

To table of contents

In Virginia, a state with high interspersion of cover types and a long history of modification of the land, aerial video proved to be a good method to collect land cover data. Aerial video allows inventory of land cover at a scale similar to the TM scene. Video transects are suitable to determine land cover type distributions in inaccessible areas, and the swath of vegetation visible provides important context that point data does not contain. Figure 4 shows the change in vegetation across a ridgetop, with mesic hardwoods on the northwest slope (A), yellow pines on the ridge (B), and dry site oaks and beeches on the southwest slope (C). We were unable to interpret to Alliance level in some situations, but more detailed data can be collected on the ground if needed. We will interpret to the Supertype level in the central Piedmont and eastern Coastal Plain regions to increase the interpretation rate and number of points accumulated. The points will be used to validate models of land cover types and to assess the accuracy of the land cover/land use map.
 

Figure 4. Ridge vegetation.

To table of contents

Anderson, J. R., E. E. Hardy, J. T. Roach, and R. Witmer. 1976. A land use and land cover classification for use with remote sensor data. U.S.D.I., U.S. Geological Survey Professional Paper 964. Washington, D.C. 28pp.

Graham, L. A. 1993. Airborne video for near-real-time vegetation mapping. Journal of Forestry 91:28-32.

McCombs, J., S. Klopfer, D. Morton, and J. Waldon. 1997. An alternative approach to land cover mapping in complex terrain. Pages 21-22 in GAP Analysis Program Bulletin No. 6.

Slaymaker, D. M., K. M. L. Jones, C. R. Griffin, and J. T. Finn. 1996. Mapping deciduous forests in southern New England using aerial videography and hyper-clustered multi-temporal Landsat TM imagery. pp. 87-101 in Gap analysis: a landscape approach to biodiversity planning. J.M. Scott, T.H. Tear, and F.W. Davis, eds. American Society for Photogrammetry and Remote Sensing, Bethesda, MD.

Society of American Foresters. 1980. Forest cover types of the United States and Canada. F.H. Eyre, ed. Society of American Foresters, Washington, D.C. 148pp.

Woodward, S. L., and R. L. Hoffmann. 1991. The nature of Virginia. Pages 23-48 in K. Terwilliger ed. Virginia's endangered species: proceedings of a symposium. McDonald and Woodward Publishing Company, Blacksburg, VA.

To table of contents
Comments or Questions?  Contact Stacy McNulty at smcnulty@vt.edu