AirSpex 2007

6.2. Data analysis II— More classes and Bayes vs. minimum distance classification


After we had finished the classification with the four “obvious” classes; snow, ocean, buildings and roads, we wanted to compare the results obtained by using the two different classification methods provided by the Image Calculator; Bayes and Minimum Distance classification.


We performed this analysis using 9 classes. We added 5 new classes to the ones used in the previous analysis, distinguishing between 5 different colors for the roofs of the houses in Longyearbyen; Black roofs, Red roofs, Blue roofs, Grey roofs and Dark grey roofs. Figure 6.3.1 displays the model areas for each of the classes. We deliberately chose the additional classes to be classes which we did not expect to be as well separated as the 4 classes in the previous analysis.


Figure 6.3.1: Model areas for the classes snow (white), black roofs (black), red roofs (red), blue roofs (dark blue), grey roofs (grey), water (turquoise), streets (orange), dark grey roofs (green).


We used the last 15 of our images in this analysis, instead of the first 15, which we used in the previous analysis (Image Calculator can only handle 15 images at a time).


Bayes classification


Bayes classification uses Bayes formula, with some modifications (amongst others, it assumes a Gaussian probability density function), in order to classify images.


Figure 6.3.2: Classified image using the Image Calculator with Bayes classification. The chosen classes are shown in the figure.


Figure 6.3.2 shows the result using Bayes classification. The result appears to be fairly well matching with the actual image, also for the roofs. UNIS’ dark grey roof is quite well classified, and so is Lompensenteret’s red roof (with some snow on the roof) and Næringsbygget’s blue roof. When it comes to all the buildings that appear to be classified with partly grey and partly blue roofs (like, for instance, the SAS Hotel, the kindergarden and Svalbardbutikken), the result is not unexpected. These buildings were never selected as model areas, and when two of the team members discussed their perceptions of the colors of these roofs, they disagreed as to whether they were if fact blue or grey. Hence, the result, that these buildings are classified as both grey and blue roofs, might just appear to be fairly reasonable.

The same goes for the buildings with dark grey and black roofs (like, for instance, the post office or the terrace houses to the right in the image).


We also see that smaller parts (in general, the edges) of most of the buildings are classified as road. For some of the buildings, there is probably some road surrounding the building (like, for instance, the road surrounding UNIS and the parking lot area outside Lompensenteret), which may partly explain this. Other than that, we probably have to look to the separation of the classes for the road and the houses, in order to explain this feature.



Figure 6.3.3: Spectral fingerprint of each class (wavelength vs. intensity).


Figure 6.3.3 illustrates the mean spectra of each class for the 15 last images in selected wavelengths. As with the previous analysis, we can clearly see that the snow is well separated from the other classes. As expected, it is reflecting with the highest intensity of all the classes, over the visible range of the spectra. We also expect the water to absorb the most light, due to the low albedo of water. Figure 6.3.3 shows that this in fact the case. When it comes to the different colored roofs and the road, the separation between the classes is not as obvious. This supports the results of the classification seen in Figure 6.3.2, where for instance parts of houses are wrongly classified as road, and the colors of some of the houses are not unambiguously decided.


Minimum distance classification


Minimum Distance classification is similar to Bayes classifier, but it is simpler and demands less computer power (information about the covariance matrix is skipped).

Figure 6.3.4: Classified image using the Image Calculator with Minimum Distance classification. The chosen classes are shown in the figure.


Figure 6.3.4 shows the result using the Minimum Distance classification in the Image Calculator. The result of this classification also appears to be fairly well matching the actual image, but we encounter some more problems regarding the “central” classes of the roads and the roofs of the building. In particular, we see that much of the road is now classified as grey roofs. We also see that several of the buildings are now wrongly classified as road. These results, which appear to be poorer than the results obtained with Bayes classification, may  said to be expected, as the Bayesian classifier is more accurate .