Hyperparameter estimation and plug-in kernel density estimates for maximum a posteriori land-cover classification with multiband satellite data

Classifying land cover via satellite imagery is an important problem in geographical studies. This paper presents a maximum a posteriori (MAP) land-cover classifier for multiband satellite data. The method uses the Markov random field model. The MAP estimation is carried out via iterated conditional modes. The reflectivities are transformed via principal components to overcome correlation and skewness, and then the distributions of the components are estimated with kernel densities. To make use of a large amount of past work in the area, the prior distribution is selected from the US Geological Survey’s land-cover database. Prior hyperparameters are estimated by equating some of their differences with the logarithms of ratios of relative frequencies from the land-cover survey, and then estimated via least squares and a bagging-type procedure. The resulting classifier produces a smooth image and captures detailed features of the study area, including roads.