FOREST SPECIES CLASSIFICATION AND TREE CROWN DELINEATION USING QUICKBIRD IMAGERY

摘要

Efficient forest management requires detailed knowledge of forest stands, including species information and individual tree parameters. Remote sensing data are increasingly being used to investigate forest classification at both coarse and fine levels. In this paper, we first examined the capability of QuickBird multispectral imagery for species level forest classification using eCognition software and a rule-based classification with the assistance of ancillary topographic data. We then applied a local maximum filter and watershed segmentation algorithm to perform tree identification and tree crown delineation using the QuickBird panchromatic band. The QuickBird imagery used in the study was acquired over Heiberg Memorial Forest in Tully, New York on 9 August 2004. For the species classification, image objects were extracted as classification units with a multi-resolution segmentation algorithm in the eCognition software. Fifty-three features including spectral metrics, texture, elevation features, and geometric features were calculated for each image object. Existing ground reference records were used for training and evaluation using the See5 data mining tool. Classification trees were built and results were evaluated using a cross-validation approach. The overall accuracy of the results was 76percent, while the lowest producer's accuracy (27percent) suggested confusions exist. Forest species classification was followed by individual tree delineation. We examined the performance of an existing algorithm by visually comparing results in three different scenarios: Emerge aerial imagery for a coniferous-dominant area, and QuickBird satellite panchromatic images over a coniferous-dominant area and over a deciduous forest stand. Preliminary results showed the tree identification and tree crown delineation algorithms were most applicable for coniferous trees in the Emerge image. Tree-top identification performance was a critical factor that influenced the accuracy of tree crown delineation.