Hierarchical Classification of Vegetation Cover Using Decision Tree Method

LANDFIRE is a large interagency project that supports national fire management and fuels treatment activities. A primary data product of LANDFIRE is a detailed vegetation type classification. This study describes how independent data from Landsat imagery, topographic data, biophysical data layers, and dependent field-collected data can be used in a hierarchical mapping approach to improve vegetation classification in the central Rockies of Montana and north-central Idaho. Field reference data included 5,350 plots that represented 14 different forest types and 1,788 plots that represented 15 different shrub types. Ninety percent of these data were used to train the classifier, and the remaining were withheld for accuracy assessment. A two-level hierarchical classification system was used to identify vegetation types of forest and shrub. A decision tree model was constructed at the first level to separate training data into distinct vegetation subtype categories based on several criteria, such as structure and leaf type. Separate decision tree models were created at the second level to further divide each subtype into the final vegetation types. Pine, spruce-fir, broadleaf deciduous forest, and other forest were identified as first-level categories of forest. Sagebrush, mountain shrub, desert shrub, and riparian shrub were identified as first-level categories of shrubs. Conventional flat, non-hierarchical classifications of forest and shrub were compared with the hierarchical approach-based classifications. This comparison showed that the hierarchical approach improved shrub classification (from 69.1 to 73.6 percent), but had no obvious advantages for mapping forest types (from 64.7 to 64.1 percent). The differences may be explained in part because species’ habitats differ widely among shrub subtypes, while there are few differences among forest subtypes. This result suggests that hierarchical classification has the potential to improve vegetation cover classification if there are significant environmental differences among vegetation cover types.