Integrating spatial and spectral information for automatic feature identification in high resolution remotely sensed images [electronic resource] /

by Lee, Jong Yeol.

Integrating Spatial and Spectral Information for Automatic Feature Identification in High Resolution Remotely Sensed Images Jong Yeol Lee This research used image objects, instead of pixels, as the basic unit of analysis in high-resolution imagery. Thus, not only spectral radiance and texture were used in the analysis, but also spatial context. Furthermore, the automated identification of attributed objects is potentially useful for integrating remote sensing with a vector-based GIS. A study area in Morgantown, WV was chosen as a site for the development and testing of automated feature extraction methods with high-resolution data. In the first stage of the analysis, edges were identified using texture. Experiments with simulated data indicated that a linear operator identified curved and sharp edges more accurately than square shaped operators. Areas with edges that formed a closed boundary were used to delineate sub-patches. In the region growing step, the similarities of all adjacent subpatches were examined using a multivariate Hotelling T2 test that draws on the classes’ covariance matrices. Sub-patches that were not sufficiently dissimilar were merged to form image patches. Patches were then classified into seven classes: Building, Road, Forest, Lawn, Shadowed Vegetation, Water, and Shadow. Six classification methods were compared: the pixel-based ISODATA and maximum likelihood approaches, field-based ECHO, and region based maximum likelihood using patch means, a divergence index, and patch probability density functions (pdfs). Classification with the divergence index showed the lowest accuracy, a kappa index of 0.254. The highest accuracy, 0.783, was obtained from classification using the patch pdf. This classification also produced a visually pleasing product, with well-delineated objects and without the distracting salt-and-pepper effect of isolated misclassified pixels. The accuracies of classification with patch mean, pixel based maximum likelihood, ISODATA and ECHO were 0.735, 0.687, 0.610, and 0.605, respectively. Spatial context was used to generate aggregate land cover information. An Urbanized Rate Index, defined based on the percentage of Building and Road area within a local window, was used to segment the image. Five summary landcover classes were identified from the Urbanized Rate segmentation and the image object classification: High Urbanized Rate and large building sizes, Intermediate Urbanized Rate and intermediate building sizes, Low urbanized rate and small building sizes, Forest, and Water.