Because file sizes and, therefore, processing times were quite large, the Old Forge 7 1/2' quadrangle was chosen to develop and test image processing techniques and methods. Old Forge is in the center of the southern scene and upland land cover types expected to be found in the study area are well-represented. An initial perusal of the air photos (1:58000 May 12, 1986 color infrared transparencies) produced field check sites. Detailed notes of upland land cover were made on a 7 1/2' X 15' 1:25000 USGS quadrangle during a field trip designed specifically for this phase of the project. A transparency was also made that fit over the air photos and indicated wetland areas. Upland cover types were sketched on this transparency to aid in signature and classification evaluation.

The Old Forge test area was subset from both the wetland GIS file and from the May and July LANDSAT imagery. In addition, a scanned NYS planimetric map of the Old Forge quadrangle in TIFF format (NYSDEC, 1992) was remade into an ERDAS image file with Image Alchemy and was subsequently brought into ArcView for registration. This latter file was used to help develop signatures in residential areas.

Because so much land cover information was known about the Old Forge quadrangle, a major concentration was directed on developing supervised classification techniques for the area. In total, 44 Old Forge GIS files, 29 image files (including the planimetric file), and 45 individual band calculations were created to evaluate signature and classification options.

Signature training polygons were digitized on the computer screen. Imagery was shown at 3X magnification and three to four band combinations were used with both the May and July imagery to highlight and accentuate various signatures. Signature statistics were created from training polygons with SIGEXT and viewed with SIGMAN (Table IIA.3).

The initial signature polygons were developed based upon field notes on the topographic map and examination of aerial photographs. Both Conifer and Deciduous forest cover types were taken from the center of relatively pure stands. Additional signature polygons were developed based upon comparisons of the scanned planimetric map, aerial photography, and field notes in an effort to discover the separability of Urban Open, Commercial, and Mixed Urban/Vegetation categories. After initial perusals of the signatures, additional Conifer, Cloud, and Cloud Shadow samples were taken.

Within-class signatures were examined with ELLIPSE (Figure IIA.2). Samples that were distinctly different from others of a similar category were removed as were samples that encompassed too broad a distribution of digital numbers. Widely different signature statistics must be considered to represent poor samples in within-class comparisons.

ERDAS SIGNATURE ELLIPSE PLOT
	Figure IIA.2. Within-class signatures shown with ELLIPSE. Note the strong overlap between the two signature samples. These two samples were eventually merged with other samples to form the training signature Deciduous. Axes represent the digital numbers of each band while the scatterplot shows the distribution of digital numbers.

Signatures were merged into classes, and between-class separability was evaluated with ELLIPSE (Figure IIA.3). While some overlap of signature statistics is expected, particularly with such a simple representation of distribution as ELLIPSE, excessive overlap suggests inseparable classes. An initial signature file with the following classes was created: Deciduous, Mixed, Conifer, Open, Barren, Urban Open (non-vegetated urban), Urban Mixed (residential and forest), Cloud, and Cloud Shadow. It seemed important to attempt a separation of urban land covers even though ELLIPSE did not support such separability. CMATRIX, which classifies using the signature samples as the data set (Table IIA.4), was used with the May imagery and the July imagery with each of the three primary classifiers: Maximum Likelihood, Minimum Distance, and Mahalanobis. The Maximum Likelihood classifier demonstrated the best overall discrimination between classes.

ERDAS SIGNATURE ELLIPSE PLOT
	Figure IIA.3. Between-class signatures shown with ELLIPSE. Deciduous and Conifer are distinct, illustrating excellent between-class separability. However, Shrub and Deciduous show strong overlap in bands 1 and 4. This overlap was consistent with all band comparisons, indicating that Shrub and Deciduous were not able to be separated into distinct classes.

As a result of ELLIPSE signature analysis, the Commercial class was dropped because all samples demonstrated huge statistical variations. All non-forested areas, such as residential open, residential vegetated, residential trailer park, commercial/industrial, open fields, airports, ski slopes, golf courses, gravel pits, and shrubby agricultural were combined into open categories and were further subdivided into Open and Open with Vegetation. While Barren (such as bare rock and gravel pits) was a desirable category, the available sample areas were very small and the class tended to strongly overlap urban non-vegetated.

A fourteen-band image file was created for Old Forge with all bands from both the May and July imagery and the Old Forge subset of the wetland GIS file was used to mask out wetlands. An initial classification was performed on the 14-band (all LANDSAT bands from the southern May and July scenes) and two 7-band (May all bands, July all bands) images. Signatures were evaluated both on-screen and with THRESH and additional signatures samples taken. Different a priori values were also tested. Some signatures in particular showed strong banding on the screen display. Consequently, individual bands in each scene were evaluated for striping (Table IIA.5). Although the thermal band (band 6) appeared to be very helpful in classifying urban areas and clouds, it created strong banding in all classifications and was removed from analyses. Numerous standard image processing band manipulations were tried and evaluated (Table IIA.6). Some band combinations demonstrated interesting and bizarre colors that did little to increase knowledge of the sample study area.