Segmentation, object-oriented applications for remote sensing land cover and land use classification

Multiscale segmentation, object-oriented methods in remote sensing have predominantly focused on urban applications using very fine resolution imagery. This dissertation explores three distinct but methodologically related remote sensing applications of multiscale segmentation, object-oriented classification using 30 m Landsat data. The first article reveals that object-oriented methods can achieve high classification accuracy for spectrally indistinct classes, even when forced to utilize non-ideal datasets such as hazy Landsat imagery and the "research grade" ASTER DEM. By incorporating spatial metrics, and exploiting elevational characteristics, seasonal wetlands can be differentiated from spectrally inseparable anthropogenically modified land use and from the upland, mixed tropical forest with high regional and local accuracies. The second article proposes and tests an object-oriented, target-constrained method for mangrove-specific change detection. By integrating pixel-based matched filter probability outputs with fuzzy object classification the proposed hybrid method bypass the need for exhaustive classification reducing classification time immensely. This method, then, has provided a means to globally assess mangrove stocks with the accuracy of object-based methods, but with the rapidity and repeatability found normally in less intensive methods. The third article demonstrates how both textural operators can be used at the object level for residential density classification with 30 m Landsat data. It was concluded that both mean GLCM and local Moran's I spatial statistics should be considered for the classification of residential density with the caveat that their utility is class-dependent. Object level usage of Moran's I was found to be able to be better differentiate high density land use classes while mean GLCM texture was indicated to be superior for separating low density land use and land cover. These applications demonstrate the utility of multiscale segmentation, object-oriented methods for a diverse array of environmental applications concerning land cover and land use classification.