The subject of this thesis is methods for classifying land using satellite images, and adherent parameter estimation. A satellite image consists of a set of pixels where measurements of spectral intensities are observed. Based on these spectral intensities, each pixel is assigned a class. The classification methods considered in this thesis are based on Bayesian decision theory.
Accounting for spatial and temporal dependence is of importance for classifying land. One such classification method is an autocorrelation method where the observed intensities are assumed to consist of the true intensities and some autocorrelated noise. In order to include temporal dependence, the spatial autocorrelation methods is extended to comprise information about temporal neighbours.
For applying classification methods it is necessary to estimate adherent parameters. The spatial autocorrelation parameter relies on the noise components, which are unobservable. An autocorrelation estimator based on Maximum-Likelihood estimates of autocovariances is introduced. This estimator is based on components that are differences between intensities from pixels taken on two different occasions over the same area. By doing so, this problem is easier to handle. Asymptotic properties such as strong consistency and asymptotic normality are proved for this estimator.
An efficient implementation algorithm for the autocorrelation methods is given which is necessary given the large amount of computation required. The suggested spatio-temporal autocorrelation method and some other classification methods are applied to real Landsat TM data. The results of the classifications were evaluated using data obtained through a field inventory. The conclusion from this study was that the spatio-temporal autocorrelation method performed best.