Object-based spatial classification of forest vegetation with IKONOS imagery
MetadataShow full item record
Object-based image analysis (OBIA) is employed to classify forest types, including deciduous, evergreen and mixed forests, in a U.S. National Park unit using very high spatial resolution (VHR) IKONOS satellite imagery. This research investigates the effect of scale on segmentation quality and object-based forest type classification. Average local variance and spatial autocorrelation analyses are utilized to determine the quality of segmentation. This research also examines the effect of grey-level co-occurrence matrix (GLCM) texture measures on forest classification results. The comparison of a manual interpretation revealed that three distinct levels of segmentation quality were yielded depending on scale: over-, optimal- and under-segmentation. Over-segmentation produced larger number and smaller size of image objects (or segments) than those of manually interpreted forest stands. Under-segmentation generated the smaller number of image objects with larger average size compared with manual interpretation. On the other hand, optimal segmentation with a scale (i.e., scale parameter) of 18 generated similar image objects much resembling manually interpreted forest stands in number and average size. Based on visual assessment, image segments were similar to manually interpreted forest stands in terms of location, shape, number and average size. Statistical analyses supported these results. A graph of average local variance against segmentation scale also indicated an optimal scale of 18. According to spatial autocorrelation analysis, this research found that over- and under-segmentations were related to positive autocorrelation, while optimal segmentations achieved lower, or even negative, Moran’s I values. This research discovered that optimal segmentations achieved higher accuracy of forest type classification than over- and under-segmentation. In particular, a scale of 19 produced the highest overall classification accuracies when using only spectral bands (79 % in overall accuracy and 0.65 in Kappa). The research found that the incorporation of individual texture measures did not improve OBIA forest classification at scale 19. Instead, the use of multiple texture measures enhanced OBIA forest type classification accuracies to 83 % in overall accuracy and 0.71 in Kappa by disentangling classification confusions. OBIA with multiple GLCM texture measures are expected to be a useful approach to automatically classify forest types. In addition, OBIA will play a role in closely coupling remote sensing and GIS with its ability to create a GIS database to be utilized for further GIS analyses.