METHODOLOGY

Landscape Information Infrastructure in Pennsylvania

Statewide spring/summer coverage of Landsat Thematic Mapper (TM) data provided through the Multi-Resolution Land Characteristics Consortium (MRLC) is the foundation of the Pennsylvania-GAP landscape information infrastructure. This foundation consists of hyperclusters which are built with the ISODATA facility of ERDAS Imagine. First, every pixel in each scene is distributed directly among a set of 255 clusters, with no sampling whatsoever. Then complete bandwise signature information is compiled in conjunction with the clustering, and this is used to compute relative brightness measures for visible, infrared, and greenness.

Those brightness values permit us to construct cluster image mosaics across scene boundaries. The clusters, with their tables of averaged spectral attributes, permit us to render generalized image reconstructions—which are export-compatible with the ARC/INFO Grid facility and are free of proprietary restrictions on redistribution. Statewide cluster images will be transferred to CD-ROM as a distribution medium and made available on a cost recovery basis for production of the CD-ROMs. These cluster images preserve visual landscape pattern and are free of thematic focus.

The tables of scenewise cluster properties are kept separate from the CD-ROM on diskette, which permits the tables to be augmented as we proceed with landscape interpretations of the clusters. The first such augmentation is a text-field characterization for each cluster. Next follows cluster categorization according to a modified UNESCO classification of land use/land cover which is substantially compatible with Anderson. This is a northeastern states adaptation of physiognomy and formation levels from a provisional scheme set forth by The Nature Conservancy (TNC). Landscape interpretations of clusters are formulated photointerpretively using the suite of facilities available in ERDAS Imagine.

Floristic categorizations of forest clusters are then assembled as separate relational tables keyed to each cluster. Reference to supplemental information sources and assistance of cooperators is required in the floristic interpretation phase. The base floristic categorization will reflect Society of American Foresters cover types as a point of departure for classification of alliance types. It has been determined that spatial (patchwise) specificity comes later in the analytical scenario.

The first step toward patchwise specificity is contiguity-controlled spatial filtering to merge cluster patches less than one hectare with larger neighboring patches. Another reason for preferring ISODATA clusters is that their numbering and initiation protocols induce strong correlation between cluster number and multispectral composite brightness. Since major land use/land cover differences find expression in composite brightness, attribution criteria for spatial merger can be satisfactorily handled in terms of cluster numbers for micro-patch suppression.

After imposing a one-hectare minimum on patchwise occurrence of clusters, the clusters are next vectorized via the Vector module of Imagine. Imagine is particularly advantageous in this regard by virtue of using the same vector format as ARC/INFO and supporting interactive image-based editing of such coverages. The commonality extends to virtual identity of "Clean" and "Build" operations. The initial attributes for polygons are scene ID and cluster number. These, in turn, serve to index the relational tables of cluster properties and scene metadata.

Floristic categorization is obtained from "multiway" analysis. Categories for recognition are determined from cluster characterizations. Training sets and signatures are obtained directly from the TM image data classified at the pixel level in supervised mode. A supervised strategy is also used to label clusters by classifying the cluster's mean vectors. The map of labeled clusters and the direct supervised classification are then differenced in terms of category numbers. Where the difference map is zero, there is local agreement between cluster-based classification and direct supervised classification. Nonzeros in the difference map indicate localities of disagreement and thus uncertainty. Overlaying the cluster-patch polygons on the difference map shows problem areas for classification. These are investigated with the help of cooperators to determine how GIS variables can be used to formulate rules of reclassification that will treat landscape settings selectively. Appropriate GIS variables are transferred by overlay as cluster-polygon attributes. Reclassification takes place on a polygon-by-polygon basis via ARC/INFO macros. Any remaining problems are resolved by direct interactive editing. Since the rules of reclassification represent elements of landscape understanding, they are saved in text form as well as the AMLs.

Following vegetation analysis, any additional site-level GIS variables required by vertebrate habitat models are also transferred by overlay as attributes for the respective cluster-patch polygons. What results from this phase is a one-hectare minimum database of polygonal landscape segments corresponding to patches of clusters. Since more than one cluster may occur in a particular vegetation class, polygon boundaries are not necessarily vegetation boundaries. To produce a vegetation map, the polygonal database is processed to dissolve boundaries between polygons having the same attribute. This set of "cluster-patch" polygons, then, constitutes the primary framework for the landscape information infrastructure.

Next comes a series of criterion-based polygon aggregations to a coarser scale. The scale change factors, in terms of minimum polygon size, are 5-hectare, 10-hectare, 20-hectare, and 100-hectare minimum levels. One objective in this reductive rescaling is to retain a visual semblance of landscape pattern, corresponding to views from increasing altitudes. Selected mixture and diversity attributes due to rescaling will be computed and entered in polygon attribute tables (PATs). When transferred from coarser to finer scales, such attributes provide vicinity context.

Scale generalization by polygon aggregation ensures that segments from different levels are strictly nested. When landscape interpretations are extracted from imagery of different resolutions, there is usually at least some degree of nonagreement. To overcome this lack of agreement, direct on-screen photointerpretation of TM data at a 100-hectare resolution is being developed to further differentiate between human-caused and natural vegetation types. The two classes being recognized are woody successional matrix versus anthropogenically sustained herbaceous matrix. Islands of either type less than 100 hectares are not delineated.

Boundary cutoffs in digitizing are likewise not considered significant if less than 100 hectares. This mapping speaks directly to high-level landscape fragmentation and provides a comparator for the strategy of polygon aggregation.

Each polygon data layer, representing a given scale, has a companion layer of indexing points. The layers of indexing points enable construction of polygon pyramids across scales. With the point indexing approach, pyramids can be constructed for hierarchies of imperfectly aligned polygons. It is also possible to adapt the point indexing strategy for "fuzzy" nesting.

Concurrently with Gap Analysis, a second major application of this Pennsylvania landscape information infrastructure is to formulate ecological land types and land type associations under the Bailey scheme being promoted as ECOMAP by the U.S. Forest Service. Deliberations en route to these formulations will add to the depth of landscape understanding.