Investigating New Advances in Forest Species Classification: Establishing a Baseline

Detailed forest classification provides critical information for forest managers. The potential for species level classification from remotely sensed data has been challenging in the past because of limitations of both available image data and traditional classification techniques. Such limitations may be reduced by the increased availability of higher spatial resolution imagery as well as detailed digital elevation models e.g. derived from lidar collections. This project is a multi-year effort aimed at evaluating the benefits of combining traditional image classification techniques with derivatives of active remote sensing sources such as lidar for species level forest classification. The project plans to evaluate the relative benefits of different classification schemes, considering accuracy, efficiency, and effectiveness. The project focuses on the classification of imagery for the area in and around the Heiberg Memorial Forest in Tully, New York. This requires manipulation and organization of existing forest inventory information for ground reference. This project aims to use several different classification methodologies. Traditional approaches—such as supervised classification—provide a means to generate baseline classifications of satellite imagery (Landsat) and high spatial resolution image data (Emerge airborne). Such fundamental techniques provide a structured foundation to enable comparison with other analyses. This paper focuses on generating species level classification from Landsat Enhanced Thematic Mapper Plus (ETM+) imagery. The analysis compares single date classification of Landsat imagery collected during spring, summer, and fall seasons, with the results attained from a multi-temporal classification of the three datasets. Later stages of the project will consider algorithms for topographic analyses and evaluate the relative benefits of different topographic models as supplemental information sources for forest classification. The documented utility of incorporating lidar-derived data layers is limited and more challenging. Such analysis will rely on alternative methods such as rule-based classifiers or neural networks. Both the supervised classifier and the rule-based approach utilize ground inventory data to provide reference information for classification and assessment. The results attained for the baseline classification demonstrates that it is possible to develop reasonable overall accuracy when performing a species level classification of Landsat ETM+ imagery (81%). However, the confusion shown in some of the low user’s (e.g. 20 %) and producer’s (e.g. 44 %) accuracy statistics suggest that there is substantial improvement yet to be made. Additional research work will focus on evaluating higher spatial resolution imagery as well as incorporating topographic indices derived from lidar sources. INTRODUCTION AND OBJECTIVES Overview Land cover classification provides valuable information to forest managers. Historically, such classification has focused on generalized forest classes. As higher spatial resolution data sources become more widely available, the potential for species level classification has increased significantly. While lower spatial resolution imagery provides an averaged response over a region, individual trees are visible in some high resolution imagery. This provides an opportunity to evaluate the composition of the mixed-species stands that are common in northeastern forests. However, a challenge to effectively utilizing high spatial resolution imagery is that the spectral response of an