Integrating satellite imagery and biophysical modeling to estimate rangeland health on the Little Missouri National Grasslands

Updated 06/13/2008

Mark Jensen, USDA Forest Service*
Billings, Montana

The Little Missouri National Grasslands (LMNG) of western North Dakota support the largest permitted grazing use within all lands administered by the USDA Forest Service. This fact, coupled with the need to revise current planning direction for forest and range allotments of the LMNG, necessitated that a broad-level characterization of ecosystem integrity and resource conditions be conducted across all lands within the study area (approximately 800,000 ha) in a rapid and cost-effective manner. We used existing field plot data collected for a variety of previous inventory objectives, and their maps, to develop an ecological classification map themes representing current (existing vegetation), reference conditions (potential vegetation). Their intersection allowed us to assign an ecological status rating (i.e. ecosystem integrity and resource condition) based on the degree of departure between current and reference conditions. In this paper we present a brief review of methodologies used in the development of the ecological classifications, and also illustrate their application to assessments of rangeland health through selected maps of ecological status ratings for the LMNG.

Ecological Classification

Classifications of potential vegetation environments (e.g. habitat types, range sites, ecological sites) are important to land managers because they provide a conceptual basis for determining both the resource potentials and ecological integrity of landscapes. Efficient use of potential vegetation classification in regional or subregional scale assessments of ecosystem health has been limited to date, however, largely because of traditional ecological unit mapping procedures which treat such classifications as ancillary information in the map unit description. Accordingly, it is difficult, if not impossible, to describe the precise location, patch size, and spatial arrangement of potential vegetation environments from most traditional ecological unit maps. Recent advances in remote sensing, geographic information systems (GIS), terrain modeling and climate interpolation can facilitate the direct mapping of potential vegetation through a prediction process based on gradient analysis methodologies and ecological niche theory.

We used a 30m raster-based map of four grassland, five shrubland and six woodland habitat types across the LMNG to develop a predictive vegetation map. In developing this potential vegetation map we used discriminant analysis based on six primary GIS themes: geoclimatic subsections and remotely sensed existing vegetation lifeforms and interpolated climate information, which were used as independent (predictor) variables in model construction; and 1,296 field plots with known habitat type membership which were used as dependent variables in model construction. Classification accuracy of the resultant habitat type map was assessed by a jackknife discriminant analysis procedure; values ranged from 54 to 77 percent in grasslands, 62 to 100 percent in shrublands, and 70 to 100 percent in woodlands dependent on geoclimatic subsection setting. We also developed techniques for generalizing 30m habitat type data to appropriate ecological unit maps (e.g. landtype associations) for use in ecosystem health assessments and land use planning.

Spatial patterns as an indicator of rangeland health

The size, shape and spatial arrangement of relatively fine-scale patterns are important indicators of health in rangelands dominated by grassland. Maintaining this fine-grained patterns can be a challenge when mapping land cover in these environments. We used potential vegetation in tandem with satellite imagery to develop rangeland vegetation maps. Our goals were to improve pattern delineation, classification accuracy, and hierarchical level of classification. We addressed pattern delineation through image segmentation, developing a multi-pass process for aggregating pixels into raster polygons. Here, potential vegetation was used to seed an iterative, unsupervised classification of Landsat Thematic Mapper imagery. Landscape structure was extracted from the imagery using a boundary detection filter, then used to constrain a merging program that aggregated pixels from the unsupervised classification into polygons. After polygons were delineated, they underwent supervised classification, where they were assigned cover type labels based on 2,617 ground-truth plots. Cover types were then stratified by potential vegetation, and the utility of this refinement for ecological assessments was explored.

For the study area, the method delineated 382,121 patches with a mean size of 4.07 ha. We compared this method with two similar approaches, and found that it better represented landscape pattern. Variation between polygons was maximized, whereas that within polygons was minimized. Important, fine-scale landscape features like riparian areas and woody draws were maintained. Furthermore, available ground-truth data were used more efficiently in supervised classification. In all, 32 land cover types were mapped; mean point-based accuracy was 74.4%, a differential improvement of nearly 20% as compared to an earlier classification. Improvements in hierarchical classification level also were observed: when grassland cover types (4 of the 32 above) were stratified by potential vegetation, the environmental range of each cover type appeared to narrow, bringing to light differences in characteristics like graminoid production.

We believe this type of approach can offer significant improvements both in accuracy and cost-effectiveness when attempting to assess rangeland health in large areas with limited data or data that has previously been collected for other purposes.

< Back to Agenda