Forest terrain feature characterization using multi-sensor neural image fusion and feature extraction methods

Although the processing of multi-spectral imagery from earth observation satellites has been effectively used for classification of many types of land cover, forest classification has generally been limited to broad categories such as deciduous or coniferous. The ability to identify individual forest species using widely available remotely-sensed data would be beneficial for many forestry applications. Recent studies suggest that the combination of imagery from satellites with different spectral, spatial, and temporal information may improve classification performance. This dissertation discusses the results of new biologically-based neural image fusion and feature extraction research aimed at deriving additional information from existing multi-spectral and multi-sensor imagery to improve forest classification performance. For this investigation multi-season Landsat and Radarsat imagery of the Heiberg Memorial Forest in central New York State, along with digital elevation data, was processed using an opponent-color image fusion and data mining technique, in conjunction with multi-scale visual texture enhancement, and the Fuzzy ARTMAP neural classifier. This approach is shown to enable identification of individual forest species with higher accuracy and fewer misclassifications than the traditional spectral-only maximum likelihood classification approach. The described neural image fusion approach could be readily extended to include other types of remotely sensed imagery and terrain contextual data.