New Remote Sensing Methods for Labeling Disturbance Agents in Appalachian Forests

Abstract:

Forests in the eastern United States are species rich and affected by a variety of disturbance agents such as fire, invasive insects, diseases, and storm events. Millions of hectares of forest are disturbed each year, altering the forest carbon sink and changing forest nutrient cycles. The magnitude and direction of these changes, though, can be different for different disturbance agents. For example, trees that burn in severe fire rapidly release stored carbon into the atmosphere whereas standing deadwood from insect attacks decompose slowly while atmospheric carbon is fixed in regenerating vegetation. The diagnosis and attribution of these processes require accurate and reliable estimates of the extent and frequency of different disturbance agents. Here, a new method is presented that classifies disturbance events identified using time-series analysis of Landsat TM imagery. The method exploits information about changes in the canopy heterogeneity as measured by several texture indices within forest patches. Classifiers were trained using data from the US Forest Service Aerial Detection Surveys and currently differentiate between fires, southern pine beetle, gypsy moth, hemlock woolly adelgid, beech bark disease, anthracnose, and storm events. In addition, the classifier returns a value of 'uncertain' when it is unable to make a clear determination, which is currently approximately 10% of identified disturbances. Classification accuracy for the remainder is 81%, though is variable between agents. For example, the classifier performs well in identifying southern pine beetle and gypsy moth affected areas, but poorly in identifying storms. Reliabilities are similar to accuracies for each agent. The results presented are the first yearly, regional-scale estimates of forest disturbance partitioned by disturbance agent. We find good correspondence with previously described patterns of disturbance and distribution, including direct observational evidence of their predicted periodicities over entire ecoregions. Such estimates are vital for forest monitoring and to better understand the role of the dynamic forest carbon sink in order to reduce uncertainty in atmospheric carbon models. Future work must focus on the inclusion of direct anthropogenic changes such as harvest and urbanization.