#39: Statistical modeling for improved land cover classification

Y. Zhao, X. Zhou, K. Palaniappan, and X. Zhuang

SPIE Battlespace Digitization and Network-Centric Warfare II, Volume 4741, pgs. 296--304, 2002

gis, classification, segmentation, data mining, machine learning, remote sensing

PlainText, Bibtex, PDF, DOI, Google Scholar


Novel statistical modeling and training techniques are proposed for improving classification accuracy of land cover data acquired by LandSat Thermatic Mapper (TM). The proposed modeling techniques consist of joint modeling of spectral feature distributions among neighboring pixels and partial modeling of spectral correlations across TM sensor bands with a set of semi-tied covariance matrices in Gaussian mixture densities (GMD). The GMD parameters and semi-tied transformation matrices are first estimated by an iterative maximum likelihood estimation algorithm of Expectation- Maximization, and the parameters are next tuned by a minimum classification error training algorithm to enhance the discriminative power of the statistical classifiers. Compared with a previously proposed single-pixel based Gaussian mixture density classifier, the proposed techniques significantly improved the overall classification accuracy on eight land cover classes from imagery data of Missouri state.