Rule-Based Image Interpretation

Alexa J. McKerrow and Steven G.Williams

North Carolina State University, Raleigh

The North Carolina GAP Project is using a decision rule process for mapping vegetation in the coastal plain. Building on the methodology developed by Slaymaker et al. (1996), we have advanced its application incorporating a rule image. As opposed to directly "deciding" pixels into a given land cover class, the output image is created using the rule numbers themselves. This results in a "rule image" that provides a useful intermediate step, allowing the analyst to check for logic errors by determining the amount and spatial distribution of the pixels classified by a suite of individual rules. This process of writing, applying, and verifying the rules can be repeated until the majority of pixels are coded. Once the quality control check is complete, the rule image is simply recoded to create the land cover map. Neighborhood majority functions are then used to code any remaining unclassified pixels.

The rule image makes it possible to return to an intermediate stage for cover classes requiring refinement. By isolating pixels selected by a single rule, the information used in that rule is maintained, and additional information can be added to refine the classification. For example, we can return to a specific rule that decides longleaf pine woodland and build on the unique spectral, soil, and wetland information that was used to isolate that subset of pixels. If we output the land cover class directly, the range-unique attribute combinations used to label that class are lost.

Methods

Ground Truth Data Acquisition

Four primary sources of ground truth information were collected for the decision rule mapping process. These sources include aerial videography transect data, ground samples obtained during videography training field work, the North Carolina Natural Heritage Program’s plant community occurrence database, and ground samples collected by the North Carolina Division of Coastal Management. Aerial videography transects were flown in the spring and fall of 1996 and again in the spring of 1998. The video equipment is designed to collect both wide-angle and zoom video. A time-code stamp on each frame ties it to a simultaneously collected GPS location (Slaymaker et al. 1996).

In the lab, video transects are used in the selection of sites to be visited on the ground. During the interpretation phase, if we see vegetation that is difficult to interpret, new sites are selected for field visits to verify the interpretation. Using the sources listed above, approximately 20,000 points were gathered for the coastal plain of North Carolina. The majority of these (86%) came from aerial videography interpretation. Point labels were assigned based on National Vegetation Classification alliances where possible (Weakley et al. 1998). Certain labels (small stream swamp, low pocosin) contain several alliances that could not be distinguished at the resolution of the aerial video. As new labels are added to the list, a crosswalk to alliance list is updated. One fourth of the interpreted points for each vegetation class have been reserved for a final assessment over the entire ecoregion. The assessment will be conducted after refinements are made, based on review, and after merging across Landsat TM scene boundaries has been completed.

Data Sets Used in the Vegetation Classification

Four data sources were used for mapping the coastal plain: a) U.S. Fish and Wildlife Service’s National Wetland Inventory (NWI), b) Natural Resources Conservation Service’s county soil surveys, c) Landsat Thematic Mapper imagery from the Multi-Resolution Land Characterization Consortium’s (MRLC) nationwide purchase, and d) the southeast regional land cover developed by EROS Data Center (EDC; Version 98-03, Hughes STX Corporation under contract to USGS). The EDC regional land cover is used as a masking tool. All pixels classed as urban, water, agriculture, pasture, other grasses, or barren are currently left unchanged by our work. The pixels corresponding to those that had been mapped as natural vegetation are used in an unsupervised clustering of the raw data to get 80 clusters.

Decision Rules Process

The decision rule process involves developing a database from the interpreted points which includes the values of each of the different data layers for each location. Specifically, in ERDAS Imagine, we use the "pixel to ASCII" function to obtain a file with the values from the soil surveys, NWI, and clustered image that occurs at each of the X,Y locations in the interpreted points. This database is then used to develop a series of IF-THEN statements for each of the 80 clusters. In Imagine, the rules are applied sequentially, and once a pixel has been classified using a conditional statement, it is not considered in subsequent rules. To mimic this while writing the rules, we analyze the data cluster by cluster and iteratively subset the points based on the combination of the ancillary information used to develop each rule. For example, in cluster 25, the majority of the points that occurred on Croatan soil and had a NWI code of saturated palustrine needle-leafed evergreen forest were labeled pond pine woodland. A rule would be written based on that combination of ancillary information. The points decided by that rule are then removed from the data set, and the process is repeated for the remaining points in the cluster.

Once the rules are written, a crosswalk between rule number and land cover is developed, and the rules are incorporated into an Imagine model. As mentioned above, we send the rule number to the output image which can then be queried using our interpreted points. If the pixels are decided as we expect, then the rule numbers can be recoded to create the vegetation map. If, however, we find mismatches between the land cover associated with the rule at that location and the interpreted points, we can readily trace the errors back to their origin. We can determine if the errors are due to true confusion between classes that cannot be overcome given the data we have or caused by typographical mistakes or errors in the logic of a rule. Once recoded to create the vegetation image, the regional land cover classes that were masked out are merged back with the vegetation classes. Class-specific majority filters are used to code previously unclassified pixels.

Review and Final Evaluation

The draft vegetation map for the southeast coastal plain is now available for review. In general, the comments we have received to date have confirmed limitations we had previously recognized. We know, for example, that we have not separated mesic from xeric longleaf pine woodlands. In this case, we will return to the rule image and look more closely at the spectral and soil data associated with the longleaf rules. We are retaining 25% of the interpreted points for a final assessment after the entire ecoregion is merged. In the meantime, we are working with a comparison of the interpreted vs. decided points in order to determine the level of confusion between the mapped classes. While we have reserved points for an assessment of the vegetation cover within the state, we would like to incorporate a systematic assessment of all land cover classes in both the final map and the original regional land cover.

Literature Cited

Slaymaker, D.M., K.M.L. Jones, C.R. Griffin, and J.T. Finn. 1996. Mapping deciduous forests in southern New England using aerial videography and hyperclustered multi-temporal Landsat TM imagery. Pages 87-101 in J.M. Scott, T.H. Tear, and F.W. Davis, editors. Gap Analysis: A landscape approach to biodiversity planning. American Society of Photogrammetry and Remote Sensing, Bethesda, Maryland. 320 pp.

Weakley, A.S., K.D. Patterson, S. Landaal, M. Pyne, et al., compilers. 1998. International classification of ecological communities: Terrestrial vegetation of the Southeastern United States. Working Draft of March 1998. The Nature Conservancy, Southeast Regional Office, Southern Conservation Science Department, Community Ecology Group, Chapel Hill, North Carolina. 689 pp.