Page 19 - Interp Book

The decision tree classification approach requires a training data set for each landcover
class mapped, and we used both field-collected and photo interpreted information. Air
photo interpretation used leaf-on photos. Most photo-interpreted training points were
generated via (1) generating a random grid of sample sites across the area, (2) zooming
to those locations at 1:6,000 resolution, and (3) circumscribing visually homogeneous
vegetation and assigning those points a landcover type. We checked all ground-
collected data using air photos and eliminated data points that appeared to fall within
mixed landcover based on expert judgment. In most cases, point data were double-
checked by a second worker using leaf-on photography to ensure that the correct
landcover type had been assigned to each point.

The decision tree classification process assigns pixels to landcover classes using the
statistical relationship between the training data and the satellite imagery and ancillary
data of a given area. All decision tree classifications were run using a 30 m spatial
resolution, which is the native spatial resolution for the Landsat Thematic Mapper
imagery. The classification procedure was implemented multiple times, using different
combinations of data, in an effort to maximize classification accuracy. Additional points
were often required when areas of a known landcover type were consistently missed by
the decision tree process. In those cases, staff inspected the high resolution aerial
photography and identified additional sample points of the necessary landcover type.
This process took advantage of staff ecological expertise and their experience
identifying the landcover types of Oklahoma. We generated more than 20 different
classification results.

Ecological System (Current Vegetation) Classification and Mapping: Image Object
Generation, Attribution, and Modeling
Image Object Generation and Attribution with Landcover. A one hectare minimum
mapping unit (MMU) was specified for this project. To ensure that the MMU was
achieved, a post hoc process was implemented using image objects generated with the
eCognition Developer software (Figure 3). Image objects were generated from the first
principle component of a NAIP image county mosaic that had been re-sampled to a 10
m spatial resolution. This procedure was run for each county mosaic. Some counties
needed to be divided into multiple pieces because they were too large to be processed
individually. This process produced a shapefile containing polygons that represented
homogeneous units (relative to the 10 m PCA result for each county mosaic). The
image objects were then used to summarize the classification resulting from the
decision tree classification procedure. The statistic of interest during the summarization
process was the mode. ArcGIS was used to determine the mode for each object. The
separate sets of image objects were then imported to a file geodatabase.

11