An investigation into the use of artificial neural networks and landsat thematic mapper imagery for vegetation classification in Southwestern British Columbia
Date
1998
Authors
White, Joanne Cheryl
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The goal of this research was the experimentation with and characterization of neural networks for the classification of vegetation from multi-spectral remotely sensed imagery. A literature survey was undertaken to provide a theoretical context for the use of neural networks in remote sensing image classification and to determine the extent of previous research on this topic. The research described in this thesis combines the spatial analysis capabilities of a geographical information system with advanced multispectral image analysis procedures. Existing forest inventory information is utilized along with digital elevation data to provide training data and ancillary information useful to vegetation classification. Neural network classification algorithms were used to pre-stratify a Landsat Thematic Mapper image into categories corresponding to five levels of the new hierarchical vegetation inventory system being implemented in British Columbia. Emphasis was placed on the configuration of the neural network and on comparisons with conventional maximum likelihood estimates. Limitations to the success of this approach included the restricted availability of reliable training and reference data for all of the vegetation classes as well as the disparity between spectral classes found in the image data and the desired informational classes of the classification hierarchy.