Abstract
The work presented in this thesis is concerned with the estimation of snow cover area based on multi-temporal Radarsat images from Heimdalen in Jotunheimen, Norway. An analysis of the temporal development of the backscattering coefficient for the multi-temporal images was performed. This has been considered in combination with the incident angle for the SAR-images and meteorological data. Different classification methods were tested. The classification methods were divided into two groups: unsupervised and supervised classification. The classification methods use spatial and temporal contextual information. This information is fused into the classification methods by using Markov random fields and Markov chains. The problem of discriminating between dry snow and bare ground was considered and two new methods for determination of an altitude line were analysed. Methods for using Markov random fields and Markov chains in combination with unsupervised classification methods are given. A comparison related to classification errors which were generated from test masks representing wet snow and bare ground was performed. The local incident angle effect on the different classification methods was also analysed. The following conclusions were reached:
The use of Markov random fields and Markov chains yielded an improvement in the classification accuracy. It is assumed that the use of spatial and temporal contextual information will yield an even better approximation of the snow cover area.
The local incident angle may have an impact on the classification result.
An analysis of the backscatter from forest areas in the SAR-images showed that it was difficult to separate wet snow and bare ground by using the backscatter coefficient. These were masked out when snow cover area was estimated.
The use of K-means as an unsupervised classification method yielded similar result as the supervised methods (Baye's classification rule). The use of Markov random fields and Markov chain combined with the K-means yielded equivalent improvement of the results as when Markov random fields and Markov chains were used with Baye's classification rule.
The use of the Nagler-algorithm yielded an overall higher classification error rate than Baye's classification rule and K-means. The poorest results of the Nagler-algorithm are probably caused by the fact that the reference image was not taken with the same imaging modes and geometries. The use of Markov random fields and Markov chains combined with Nagler-algorithm improved the results.
K-means combined with ''Markov random fields and Markov chain'' is a candidate for further development for operational use.