Author(s):
Junchang Ju* - Department of Geography, Boston University
Sucharita Gopal - Department of Geography, Boston University
Eric D Kolaczyk - Department of Mathematics and Statistics, Boston University
Abstract:
Pixel scale classification is widely used for classifying land cover and land use in remote sensing. One of its main limitations is that pixel scale classification does not use the spatial information inherent in the image. Methods that use the spatial information often rely on the use of a moving window of a fixed size and thus may not be ideal given the diverse shapes and sizes of land cover/use patches in the scene. In addition, the use of a fixed set of classes for all spatial scales may result
in large loss of categorical information for the rare classes at the coarser spatial scales. Our prior research designed a multiscale multigranular (MSMG) framework for land cover/use classification that adaptively chooses blocks of different sizes and class labels from a categorical hierarchy for different parts of the image
and balances the spatial and categorical complexity of the classification based on penalized maximum likelihood. Our prior research tested the framework only with small images of very simple scenes.
In this paper, we introduce a post-processing procedure to refine MSMG results and we validate the framework in classifying a large TM image characterized by a complex categorical hierarchy.
MSMG results are compared with a traditional method involving aggregating pixel-scale class labels. Evaluation of accuracy with a fuzzy rating system shows that the MSMG approach reduces the magnitude of classification errors. The MSMG classification captures the boundaries of land cover/use patches reasonably well simultaneously across classes of varying categorical complexity.
