#30: Optimal Bayesian classifier for land cover classification using Landsat TM data

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

IEEE Int. Geoscience and Remote Sensing Symposium (IGARSS), Volume I, pgs. 447--450, 2000

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

PlainText, Bibtex, PDF, URL, DOI, Google Scholar


An optimal Bayesian classifier using mixture distribution class models with joint learning of loss and prior probability functions is proposed for automatic land cover classification. The probability distribution for each land cover class is more realistically modeled as a population of Gaussian mixture densities. A novel two-stage learning algorithm is proposed to learn the Gaussian mixture model parameters for each land cover class and the optimal Bayesian classifier that minimizes the loss due to misclassification. In the first stage, the Gaussian mixture model parameters for a given land cover class is learned using the Expectation-Maximization algorithm. The Minimum Description Length principle is used to automatically determine the number of Gaussian components required in the mixture model without overfitting. In the second stage, the loss hnctions and the a priori probabilities are jointly learned using a multiclass perceptron algorithm. Preliminary results indicate that modeling the multispectral, multitemporal remotely sensed radiance data for land cover using a Gaussian mixture model is superior to using unimodal Gaussian distributions. Higher classification accuracies for eight typical land cover categories over one full Landsat scene in central Missouri are demonstrated.