Date of Award:
Master of Science (MS)
R. Douglas Ramsey
Monitoring and identifying the state of rangelands on a landscape scale can be a time consuming process. In this thesis, remote sensing imagery has been used to show how the process of classifying different ecological sites and states can be done on a per pixel basis for a large landscape.
Twenty-seven years' worth of remotely sensed imagery was collected, atmospherically corrected, and radiometrically normalized. Several vegetation indices were extracted from the imagery along with derivatives from a digital elevation model. Dominant vegetation components from five major ecological sites in Rich County, Utah, were chosen for study. The vegetation components were Aspen, Douglas-fir, Utah juniper, mountain big sagebrush, and Wyoming big sagebrush. Training sites were extracted from within map units with a majority of one of the five ecological sites.
A Random Forests decision tree model was developed using an attribute table populated with spectral biophysical variables derived from the training sites. The overall out-of-bag accuracy for the Random Forests model was 97.2%. The model was then applied to the predictor spectral and biophysical variables to spatially map the five major vegetation components for all of Rich County. Each vegetation class had greater than 90% accuracies except for Utah juniper at 81%. This process is further explained in chapter 2.
As a follow-on effort, we attempted to classify vegetation ecological states within a single ecological site (Wyoming big sagebrush). This was done using field data collected by previous studies as training data for all five ecological states documented for our chosen ecological site. A Maximum Likelihood classifier was applied to four years of Landsat 5 Thematic Mapper imagery to map each ecological state to pixels coincident to the map units correlated to the Wyoming big sagebrush ecological site. We used the Mahalanobis distance metric as an indicator of pixel membership to the Wyoming big sagebrush ecological site. Overall classification accuracy for the different ecological states was 64.7% for pixels with low Mahalanobis distance and less than 25% for higher distances.
Stam, Carson A., "Using Biophysical Geospatial and Remotely Sensed Data to Classify Ecological Sites and States" (2012). All Graduate Theses and Dissertations. Paper 1389.
Copyright for this work is retained by the student.