Application of Maximum Likelihood and Evidential Reasoning Classifiers for Mapping Conifer Understory

Information about the presence and spatial distribution of white spruce conifer understory within deciduous and deciduous-dominated mixed-wood stands is required for boreal mixed-wood management in Alberta. A set of forest land cover classes was created that consisted of 30 classes described by overstory stand structure and three levels of understory amount. These polygons were overlaid onto two-date, leaf-off and leaf-on, Landsat Thematic Mapper (Landsat TM) images from which random pixel samples were extracted for classification and independent validation. An iterative supervised classification algorithm combined with class aggregation rules were used to reduce the 30 classes to 16. These 16-classes were compared to results obtained with a knowledgebased, supervised evidential reasoning classifier. Similar classification accuracy results were obtained from the two classifiers with spectral data alone but accuracy increased significantly when information about stand structure was added to the evidential reasoning classifier. Species composition, height, and crown closure describes stand structure but are often represented on a nominal or ordinal scale that is not appropriate for statistical classifiers but can be used in an evidential reasoning classifier to improve classification performance. 1.0 INTRODUCTION Mixtures of aspen (Populus tremuloides Michx.) and white spruce (Picea glauca (Moench) Voss) occur frequently in the Mixedwood Section (Rowe 1972) of the western boreal forest. Within these mixtures, white spruce often occurs as understory trees in aspen and aspen-dominated mixed-wood stands, but current forest inventories often do not document their amount or distribution (Brace and Bella 1988; Lieffers et al. 1996). Information about the understory component is needed for spruce management planning (Brace and Bella 1988), and is important because of its future contribution to white spruce timber supply (Morgan 1991; Lieffers et al. 1996). Understory information is currently obtained by interpretation of leaf-off aerial photographs and field surveys that are costly and time-consuming to undertake when large areas are involved. A method involving satellite remote sensing data may provide useful information at the planning level by providing an initial stratification of the forest landscape for understory location, distribution and amount. A prerequisite to the mapping of conifer understory is a classification system to describe the structure of * Presented at the Fourth International Airborne Remote Sensing Conference and Exhibition/21 55 Canadian Symposium on Remote Sensing, Ottawa, Ontario, Canada, 21-24 June 1999.