US Fish and Wildlife Service 1996 Landcover

Landcover database derived from 2 co-registered multispectral landsat TM 5 scenes (P34R26; P33R26-27; P32R26-28; P31R26-29; P30R27-30; and P29R29-30). Geocoding and image-to-image referencing was conducted on each of the 16 pairs of images, which were georeferenced to USGS 1:24000 scale topographic map sheets. Six bands (bands 1-5 and 7) from each co-registered scene were stacked to produce a single 12-band imagine raster file. The Imagine ISODATA algorithm was applied to the 12-band image to generate 240 spectral class signatures The Imagine maximum-likelihood classifier was then applied using the resulting signatures to assign all image pixels from the 1-2-band image to 240 classes. The resulting classes were compared to ground truth information collected during summers 1996 and 1997 for the purpose of interpretation and grouping. The 240 classes were interpreted and grouped into eight landcover classes (grassland, undisturbed grass, alfalfa hayland, cropland, forest, barren, water, and riparian). A 3x3 cell majority filter was passed once over the resulting landcover image to eliminate noise. he image was then tiled into 9 separate subwindows which were each processed with the Imagine clump process to identify contiguous areas of landcover. The imagine eliminate process was then conducted on each of the nine clump subwindows to eliminate clumps smaller than 10 pixels, or five acres in the area. The generalized images were then stitched together using the Imagine mosaic process. Wetland basin information provided by the US Fish and Wildlife Service were overlaid with the resulting 8-class image to add wetland classes (temporary, seasonal, semipermanent, lake, river). Manual interpretation of the image to identify urban built up landcover was conducted by EarthSatellite Corp. These urban built-up areas were overlaid with the landcover image and combined with the barren landcover class. An accuracy assessment was conducted upon the completion of the classification process for each scene. Ground data collected on 954,865 acres was partitioned into 2 data sets, one was used as training data and the remaining 50 percent was used for assessment of classification accuracy. An assessment of accuracy was conducted both on individual scenes and on a final mosaiked image. Imagine software was used to stitch all 16 scenes together into one image using the mosaic command. Users accuracy for individual scenes for both the grassland and undisturbed grassland was used to determine the order of overlay in the mosaiked image. Classification accuracy for individual scenes ranged from 80-92% and an assessment of the mosaiked image indicated an accuracy more than 80 percent.

spatial and attribute data

code representing 1 of 16 classified landcovers

these codes are specific to resource issues in the prairies concerning nesting waterfowl. A crosswalk to NVCS has been completed and is available from the HAPET Office.

Landcover entities are identified through attributes for image cell values. A single landcover type is assigned as an attribute to each image cell. The SDTS model of spatial phenomena describes the real world as consisting of entities which are characterized by attributes which have attribute values. Imagine raster images do not explicitly use this entity-attribute-attribute value model.

Items, Codes, and Description of each code