Flexible Image Classification Using ARCGIS

The Alaska Landcover Mapping Program develops landcover products for National Park Service (NPS) lands in Alaska as part of the NPS Inventory and Monitoring Program. Our goal is to provide reliable and consistent scientific information on a Park-wide basis that can be used to assess the status and trends in condition of ecosystems within National Parks.

Alaskan Parks comprise 54,000,000 acres in 15 units throughout the state. These Parks represent a wide diversity of habitats and vegetation. Landcover mapping projects are being developed for each NPS unit. The Katmai National Park and Preserve Mapping Project was initiated in 2000 and completed in 2002. The study area included over 4,000,000 acres of rugged, remote wilderness.

Landsat TM 7 data from four satellite images were used to develop a landcover classification for Katmai National Park and Preserve. The initial spectral analysis was performed under contract by Geographic Resource Solutions, Inc. (GRS). Supervised and unsupervised techniques were used by GRS to generate a spectral database of over 1300 unique spectral signatures. These spectral signatures were then assigned vegetation cover attributes based on available field data.

Map categories were assigned to each signature based on the vegetation attributes in the spectral database using a rule set that was intended to identify the important landcover types encountered in the study area. As a result the 1300+ signatures which were found in the spectral database were grouped into 37 calculated map classes.

The spectral database with calculated landcover type was provided to NPS as a GRID geodataset. Since the design of this project stressed systematic processing methodologies and the implementation of a disaggregated spectral database product, NPS was able to implement additional rules based evaluations using ARCGIS to further refine the landcover map (See Katmai.jpg).

ARCGIS Implementation
1) Signature Evaluation
Several sources of error are generally identified during a review of map accuracy. Frequently the field data that were used in the analysis can be erroneous or incorrectly interpreted. This will result in the incorrect assignment of class type to a signature. Observer error is especially problematic in “shoulder” ranges where one map class transitions to another class. Minor variation in observer estimates of cover at these shoulder ranges can result in major changes in signature naming. In addition signatures can perform well in a limited area or self classify however extension of the signature throughout the study area can result in misclassification. This is especially true in woodland and open canopy classes where changes in understory are not reflected in the map class name but could potentially alter the reflectance characteristics of the site.

Due to restrictions in field time/money and logistical constraints it was not possible for the initial field team to visit all areas and vegetation types within the Park. Subsequent visits to the Park provided additional field reference data that were not available in time for the spectral processing.

Since the detail of the original spectral database was maintained in the GRID implementation, misclassification could be identified and isolated. Corrections were accomplished by simply adding additional attributes to the spectral database and recoding. In this manner all corrections could be systematically tracked.

Some misclassification errors were specific to regions of the park. These errors were isolated using masks and reevaluated. Cut lines between mask regions were minimized since errors were identified in the spectral database and corrections were selectively applied to individual signatures rather than an entire mapping class (See geomask.jpg).

2) Interpretation Models
Spectral classification provides an important first interpretation of satellite images according to the various “colors” of the image. However map making generally includes interpretations that rely on analyst knowledge of the study area and the context of the sites being interpreted. Interpretation is generally incorporated in the digital map making process through ancillary data or models that are used together with the results of spectral analysis.

Water Model
Spectral confusion between water, conifer, terrain shadow and wet herbaceous/moss types is common for landcover classification efforts using TM7 data in Alaska. In an effort to successfully classify the wet herbaceous areas during spectral analysis a pronounced misclassification of water was tolerated. In addition, a slope mask had been used in generating the spectral database to minimize water misclassification in shadows. The slope mask had resulted in artifacts between DEM tiles within water areas and in areas of steep shoreline. A water model was developed from the images used in the analysis to overcome these problems (See watermask.jpg).

This water model was derived from a summation of bands 4 and 5. Both of these bands are very sensitive to water absorption. A mask derived from band 4 by itself was unsatisfactory since sedimentation produced too much variability in the water threshold. A mask derived from band 5 by itself was unsatisfactory since solid water (snow and ice) was as absorptive of band 5 as liquid water. However the sum of these bands provided an index with a reliable threshold for the discrimination of liquid water bodies from the surrounding land. Shadow areas within the mask remained problematic as they could not be distinguished from water and therefore a slope mask was developed from the available DEM. Areas of the water mask that occurred within the slope mask were eliminated. The resulting water mask was merged with the classification. The Water Model illustration in the poster compares an area of the original classification with the final classification after the application of the water model.

Winter Image Model
A broad range of wet vegetation types are generally incorrectly classified as conifer types using only spectral information. A winter image for the study area was purchased and coregistered with the existing classified product. Initial evaluation of the image revealed significant shadowing from large mountains due to the low sun angle. Band 1 provided the best penetration of these shadow areas and was selected for analysis. Thresholds were evaluated in areas where known dense conifer stands were confused with wet vegetated areas. A conservative threshold was identified that was most effective in identifying conifer from wet areas and a model was generated. The Winter Image Model illustration in the poster represents an area of the original image, the winter image, the classification before the winter model and the classification after application of the model for discriminating dense conifer and conifer/broadleaf from wet areas. During evaluations it became apparent that a breakpoint could be identified that would allow areas of lower density conifer and broadleaf cover to be identified and another model was generated. The winter image models were used together for additional reclassification (See wintermask.jpg).

Terrain Shadow Model
An illumination correction was used to preprocess the satellite imagery by GRS. It appeared to have successfully minimized the impact of terrain shadow on the spectral classification however in areas of extreme topography some problems were still noted.

A mask was generated by level slicing that corresponded to dark areas in the image Slope was used to separate the steep shadow areas from other dark objects and the resulting output was used as a terrain shadow mask. Map classes within the terrain shadow mask were evaluated and confusion pixels that fell within the shadow mask were extracted from the image and reclassified based on photointerpretation. The Terrain Shadow Model section in the poster illustrates this evaluation (See shadowmask.jpg).

Summary
As a result of this work a final generalized landcover map was generated for the entire Park. However since the original, detailed spectral database is preserved along with the generalized landcover classes, additional more refined map classes can be derived as required for specific applications or areas within the Park.