Predicting spatial distribution of privet (liguestrum spp.) in South Carolina from MODIS and forest inventory plot data
Privet's aggressive competitive behavior causes environmental harm to the ecosystem by degrading species diversity and wildlife habitat. Effective control of its spread requires high-quality spatial distribution information. Our objective in this study was to evaluate the capability and reliability of using satellite imagery, ancillary, and forest inventory data to predict and map the spatial distribution of privet (Ligustrum spp); an invasive species. Two types of trials (one for forested area and the other for the entire state's area), and two datasets for each trial were used to model privet spatial distribution within forest. The first dataset in each trial has all forest plots with privet coded as present or not present, while in the second, each forest plot with privet has an assignment to one of four classes, based on the proportion of privet in the plot. Separately, each dataset was used in a decision tree classification process applied to MODIS satellite data (250-meter resolution) and ancillary data to classify and model privet spatial distribution. Classification results from the first trial (forested area) show a decrease in overall classification accuracy from 84.73 percent, when plots are coded for presence or non-presence of privet, to 78.64 percent, when each privet plot uses categorical codes based on field calls for average privet proportion present on the plot. The decrease in accuracy is the result of a combination of assigning privet classes and an insufficient number of plots in classes containing greater than 10 percent privet. Overall classification accuracy of 72.30 percent in the second trial decreased to 69.39 percent when the privet-presence class was expanded by using the field-call categories. The greatest increase in accuracy is exhibited when comparing classifications of the entire state area (69.39 percent accuracy) with the state's forested area (78.64 percent accuracy). In both trials, plots with privet were divided into four classes, based on the proportion of privet. The final output provides information on spatial distribution of privet in forest areas and at the state level. To improve model performance, there needs to be an increase in the number of forest plots containing greater than 10 percent privet cover.