Classifying and Mapping Forest Cover Types Using Ikonos Imagery in the Northeastern United States

Accurately mapping forest cover types in the northeast using remotely sensed data has proven problematic. High species diversity and spatial variability make creating training areas of spectrally “pure” classes of forest cover types difficult. The advancement of higher spatial resolution satellite sensors have historically allowed for increases in the accuracies of mapping forest cover types. These accuracies have been further increased by image processing techniques such as hybrid forms of the supervised and unsupervised classification methods. However, since the launch of very high resolution (VHR) sensors (≤4m) such as IKONOS, traditional per-pixel automated classification has not always worked well. While increases in spatial resolution create increased amounts of informational detail, this also creates higher within class spectral variability, potentially causing lower classification accuracies when solely using per-pixel classification techniques based upon spectral comparisons. By incorporating other image processing techniques such as texture into automated classification, the increased within class spectral variability inherent to very high resolution images may increase our ability to discern relatively homogeneous clusters of pixels. Segmenting images into clusters of unique texture followed by classifying those clusters based on their spectral response patterns is a potential method for both: 1) increasing the accuracy of utilizing very high resolution images for classification, and 2) increasing the accuracy of mapping forest cover types in the northeast. Preliminary results considered in the context of an index indicate that the per-segment approach to forest cover type classification yields a significantly better accuracy compared to the per-pixel approach.