Schetselaar, Ernst
4 CLASSIFICATION OF LEVEL 2-3 AREAS
Three classification experiments to predict level 2-3 areas were conducted: one employing the TM bands,
one employing the magnetic susceptibility channel and a classification on the magnetic and TM channels
combined. All the image channels were augmented with their average filtered derivatives to exploit the
spatial autocorrelation between pixels (Switzer, 1980). This method is applicable to the purpose of this
study, because the units to be classified occur on a larger scale than a single pixel. Experiments with
different classification methods revealed that the classification was not sensitive to a particular algorithm.
The best results, however, were obtained with the Maximum Likelihood Classifier in which prior
probabilities were set in proportion to the distribution of training samples over the two classes.
No unclassified pixels were allowed to map the patterns with the greatest structural continuity possible.
Attempts were made to utilize the DEM alone or in combination with TM and Magnetic channels. The
resultant classification map patterns did not readily match any of the general geology and degraded the
overall classification accuracy. The DEM was thus used only as a backdrop image to display the data for
interpretive purposes. Future investigations aim at analyzing relationships between topography and the
major classes. For example it may be possible that topographic highs could be used to isolate cover unit
klippens, as is commonly observed in other fold and thrust terrains. Unfortunately this preliminary
investigation did not reveal consistent relationships between the basement-cover contact and topography
throughout the study area. Figure 5 shows the resulting map patterns of our classification experiments.
Table 1 provides the confusion matrices obtained from the experiments.
Table-1: Classification Results Incognita Peninsula, Nunavut
Landsat TM (14 channels)
Basement Cover
Basement 86.2 % 49.0 %
Cover 13.8 % 51.0 %
Overall accuracy: 71.3 %
Magnetic susceptibility (2 channels)
Basement Cover
Basement 78.5 05 27.4 96
Cover 24.5 % 72.6 %
Overall accuracy: 74.3 %
Magnetic susceptibility and TM bands (16 channels)
Basement Cover
Basement 79.0 % 27.8 %
Cover 21.0 % 72.2 %
Overall accuracy: 76.2 %
Note: Classification results of Maximum Likelihood supervised technique applied in various band
combinations. À combined quaternary, vegetation and hydrography mask was used to exclude areas from
the classification. Total accuracy calculated by summing all correctly classified pixels and dividing by the
total unmasked pixels. Masked pixels are excluded from the calculation.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 1329