Kontekstuell klassifisering av bilder for å kartlegge akvatisk vegetasjon

Abstract

Airborne photography are used to map zones of potential aquatic vegetation. This is done by statistical classification using different approaches. We use five classes representing different vegetation and landscape zones, one being the vegetation we want to map. The color signal in the photographs are treated as multivariate normal distributions for each class. We investigate different approaches to estimate the parameter's of these distributions. One approach being to use the data from a training sample. Another approach is to use EM. Our investigations show that the first approach gives the best result. We also looked at regularized co-variances which avoids overlap between classes. To adjust the parameter in the regularization we used the error rate of our vegetation class, and tried to minimize this. This seemed to give better results. Further on we expanded the model to include MRF. This has the effect of improve the smoothness of the final map, and avoid some of the distortions. Also here we looked at different ways to estimate the parameters. Once again it looked as if the EM algorithm didn't function well in such a model. We tried also an pseudo-ML approach to estimate the parameter of the MRF-model. This seemed to give fair result, but maybe not for our specific class. It seemed more like a global result. This was confirmed when we used a simulated photo. We were also investigating how many classes we should include in the model. There was no straight answer for this. It seems like a manual approach is necessary to find the optimal classes. What we saw was that it's beneficiary to choose small sizes for the training sets. We were also looking at how these classes could be used in a overlapping photograph. Our results seems to show that this is cumbersome.