Vegetation Community Classification and Mapping of the INL Site




Shive1, J.P., A.D. Forman1, K. Aho2, J.R. Hafla1, R.D. Blew1 and                           K.T. Edwards1


1S.M. Stoller Corporation (now Gonzales-Stoller Surveillance, LLC), Environmental Surveillance, Education, and Research Program, Idaho Falls, ID

2Department of Biological Sciences, Idaho State University, Pocatello, ID


January 2011

Vegetation Community Classification and Mapping of the Idaho National Laboratory Site Report


Executive Summary

The Idaho National Laboratory (INL) Site is located in southeast Idaho and occupies 2,300 km2 (890 mi2) of sagebrush steppe. The INL Site is managed by the U.S. Department of Energy (DOE) and serves as a science-based, applied engineering national laboratory that supports the DOE missions in nuclear and energy research, science, and national defense.


The most recent vegetation mapping effort at the INL Site was almost twenty years ago and does not capture important habitat changes that have occurred since. Prior mapping efforts also lack assessments of accuracy, making it difficult to quantify uncertainty associated with habitat models. Accurate classification and mapping of vegetation communities have become increasingly important tools for conservation management.

The goal of this project was to develop an updated vegetation map defining the distribution of plant communities on the INL Site. Our specific objectives included: 1. characterize the vegetation community types present on the INL Site; 2. define the spatial distribution of those community types; and 3. conduct a quantitative accuracy assessment of the resulting map.

Objective 1 – Plant Community Classification
We completed two separate classification efforts to support this project. A preliminary classification was conducted using previously existing vegetation data from the INL Site. The purpose of the preliminary classification was to identify the range of vegetation types potentially occurring on the INL Site and to reconcile those types with the National Vegetation Classification. Results associated with the preliminary classification were used to generate a working list of plant communities and a key for field identification of those communities. The preliminary plant community list and key were also utilized to direct sampling efforts for subsequent vegetation data collection to support a final classification.

Vegetation data were collected on 314 plots for the final vegetation classification. Plots were initially selected according to a stratified random design using Geographic Information System (GIS) data layers including previous vegetation maps updated with current wildland fire boundaries. We used the preliminary plant community key to modify plot locations periodically during the field season, which ensured all potentially occurring vegetation types were adequately sampled. A quantitative final classification was completed using absolute cover by species data from each plot.

The analytical approach to classifying the vegetation cover data was a multi-step process. First, we identified the best classification model for describing the structure and pattern of species abundance and composition. Next, we determined the optimal number of clusters, or vegetation classes, within the dataset. Upon selection of the most appropriate classification, we re-evaluated several clusters within that classification to determine whether they should be further split. Finally, the classification and cluster summaries were updated to reflect additional cluster divisions.

The final classification combined with the subsequent iterations of classification refinement resulted in 26 vegetation classes for the INL Site. Of the 26 vegetation classes identified, two are wooded or woodland types, seven are shrubland types, four are shrub herbaceous types, five are dwarf shrubland or dwarf-shrub herbaceous types, five are herbaceous types, and three are semi-natural herbaceous types. Semi-natural types are generally defined as being dominated by non-native species. We classified 14 of the 26 classes at a hierarchical level comparable to an Association within the National Vegetation Classification (NVC), while the remaining 12 classes were classified at a level comparable to an Alliance. Upon completion of the final classification, we used the resulting plant community class list to identify polygons delineated through the mapping process. We also developed a dichotomous key based on the final classification to facilitate data collection and support an accuracy assessment of the final vegetation map.

Objective 2 – Vegetation Class Delineations and Mapping
In 2007, we had four-band color-infrared 16-bit orthorectified digital imagery collected at 1 m spatial resolution across the INL Site. This imagery served as the source dataset for map delineations. We also incorporated the 2004 and 2009 National Agricultural Imaging Program (NAIP) imagery to define class boundaries in the areas where clouds and cloud shadows obscured the ground, and to refine class boundaries where wildland fires burned in 2007-2008. To assist with the vegetation class delineations, we calculated two vegetation indices (i.e., the Normalized Difference Vegetation Index and the Soil-adjusted Vegetation Index), as well as a statistical texture layer (i.e., 3x3 Range) using the digital imagery. We also used ancillary GIS data layers (e.g., wildland fire boundaries, DEM, etc.) during the image delineation process.

We understood the possible limitations of automated classification methods in a semi-arid sagebrush steppe environment and conducted manual photointerpretation of digital imagery directly within a GIS. The initial draft delineations were produced through manual interpretation and digitizing at a 1:12,000 mapping scale. Occasionally, we adjusted the GIS display zoom to coarser scales (e.g., 1:24,000) where broad landscape patterns were more evident. We also considered DEM topographic contours which sometimes helped delineate class boundaries. There are five non-vegetation classes and one agricultural class we digitized at a 1:2,000 scale and included in the final map. The vegetation map data will contribute to a number of ongoing and future studies on the INL Site, and we wanted to make sure anthropogenic features are not included in the actual vegetation polygons where they could negatively impact other studies.

After we completed the draft delineations for the entire INL Site, we made numerous visits to the field to investigate the communities present on the ground. The first important observation we made in the field was the initial draft delineations captured too much detail and many times the same vegetation community extended across multiple map polygons. Another important observation we made was that the majority of mapped polygons contained multiple vegetation communities present on the ground forming multi-class complexes. We edited the draft delineation boundaries and assigned each map polygon to a vegetation class or two-class complex.

The final vegetation map contains a total of 2038 polygons, of which 1964 (96.4%) represent vegetation communities. The remaining 74 polygons (3.6%) represent non-vegetation or agriculture classes we included in the map. The smallest mapped polygon, not part of a special feature or at the edge of the INL Site, is 0.0021 km2 (0.52 acres). The largest polygon we mapped is 236.3 km2 (58,399.6 acres) located in the undisturbed interior portion of the INL Site. The mean area for all vegetation map polygons is 1.1 km2 (286.8 acres). A total of 127 vegetation map classes were produced when including all two-class complexes. Twenty-two map classes were stand-alone classes as originally defined through statistical analysis. Nearly half the INL Site area was mapped as single vegetation classes and the most common stand-alone class was the (2) Big Sagebrush Shrubland class. Of the 127 total map classes, 30 classes (23.6 %) contain only a single polygon and 76 map classes (59.8%) contain five or fewer polygons. Even though there were a large number of vegetation classes and complexes mapped, the majority of those classes are limited in frequency and distribution. The remaining 51 map classes contain the majority of mapped area on the INL Site (about 85%).

Objective 3 – Vegetation Map Accuracy Assessment
We sampled 502 validation plots in 2009 using a plot array design where five subplots collectively represented a single accuracy assessment location. The rationale for multiple subplots was an attempt to capture vegetation class variability across an extent that bridged the gap between the 1:12,000 mapping scale and the original vegetation classification scale. Each validation plot array was treated as a single validation point, but because we implemented a multiple subplot design, assigning validation plots to a vegetation class or two-class complex required the development of rule sets.

We used an error matrix to calculate accuracy metrics such as user’s/producer’s accuracy, overall accuracy and the kappa statistic. Given that we had to accommodate two-class complexes in both the map polygons and the validation plot data, we devised alternative methods for populating the error matrix. The first method was a direct comparison where if a single map class within a complex matched the ground data, regardless if the validation plot was a two-class complex, it was marked correct. The second method requires both communities in a two-class complex from the map to be present in the validation plot data.

Fuzzy set theory provides an avenue to embrace multiple class membership at a single location and allows for more meaningful interpretations of the map accuracies and errors. We wanted to minimize the subjective decision making process and selected a Bray-Curtis community similarity threshold of 0.35 to identify classes eligible for fuzzy membership (Level 4 Good Answer)

Of the 502 validation plots, 186 plots (37.1%) had all five subplots key to the same vegetation class. The majority of plots that had homogenous subplot classes were in big sagebrush dominated classes. There were 226 plots (45%) that were assigned to a single class with at least three of the subplots representing the same vegetation class. Ninety plots (17.9%) were assigned as two-class complexes.

The error matrix assessment resulted in an overall map accuracy of 70.7%, a Kappa of 0.65, and individual vegetation class accuracies varied greatly. The fuzzy error matrix assessment showed substantial improvements to overall map accuracy and also individual class accuracies. The overall map accuracy increased to 94.2% and Kappa increased to 0.93.

The vegetation accuracy assessment found highly accurate results for the overall map and also individual class accuracies for most vegetation classes. Although there has never been a quantitative evaluation of previous INL Site vegetation maps, the new map is the most detailed and likely the most accurate ever produced.

Of all the vegetation map classes, the three big sagebrush-dominated classes may be some of the most important vegetation classes on the INL Site. The two most common big sagebrush classes ([2] Big Sagebrush Shrubland and [7] Wyoming Big Sagebrush Shrubland) had the largest validation sample sizes, and were found to be very accurate with the fuzzy assessment ranging from 96%-100% for both user’s and producer’s accuracy in each class. Big sagebrush communities support sagebrush obligate species (e.g., greater sage-grouse [Centrocercus urophasianus]), many of which are declining range-wide. The ability to accurately identify the distribution of sagebrush habitat has important implications for conservation management planning and the development of predictive species models on the INL Site.