For each ground truth site corresponding image coordinates were located on the IKONOS imagery to approximate ground truth site extent as indicated by corresponding colour photography. Each colour photograph was taken from a helicopter and assigned accurate GPS coordinates. Presented coordinates are of water sampling sites visible on the photography.

Accuracy of the coordinates of original data set is estimated between 50 and 360 m. Where possible, satellite image features were matched to features visible on the photo 'near' the given coordinates. Feature sizes were estimated on both the photo and images and compared for shape and colour. Unfortunately, due to inadequate feature resolution some sampling sites could not be precisely paired with known coordinates of IKONOS imagery. Inability to pair coordinates resulted from either the presence of adjacent features of similar shape or possible inaccuracies in coordinate data values.

Any classification is not meaningful until its accuracy is assessed (Hall, 1998). Error matrix (or confusion matrix) compares on a category-by-category basis the relationship between known reference data (ground truth) and the corresponding results of an automated classification.

At the end of each classification, the resulting thematic map was compared to available ground truth data. The B/W air photographs were only of limited use because of their age and scale. Errors are identified where classification clearly missed some land features.

When the Landsat image was classified based on training areas created only in Scotty Creek basin, the classification results were inaccurate. The classifier confused many sparse coniferous forests with channel fens anddecidous forests. The shallow ponds on the channel fen were not properly classified because no such land cover classes existed in Scotty Creek basin. In the second classification, the new class “ponds” was added. The highway was also not classified, again due to lack of that type of land cover in Scotty Creek basin. After the training areas were improved (more samples from outside Scotty Creek basin), the classification improved considerably. On the colour photos, the deciduous trees appear as yellow due to change in leaf colour in the late summer. Channel fen is an elongated brown coloured feature and coniferous trees look dark green.

The sources of error in this classification can be attributed to several factors. In many cases, the reflectance of one feature could be similar to the reflectance of another feature, resulting in confusion. The similarity in reflectances could be the result of similar background components and variations in tree density. Error could also be a result of spectral mixing of various features that fall within a 30 meter pixel (Yamagata, 1996a). This is particularly evident in Landsat pixels as compared to IKONOS pixels. The Landsat Thematic Mapper sensor detects the combined reflected light from the 30 m by 30 m area. As a consequence, if different landscapes are present, for example wetlands and forest patches, then the resulting pixel might look as a “very green wetland” or “very wet forest”, and in general create confusion during classification. In this way, many pixels will be improperly classified.

Since the IKONOS raster has greater spatial resolution (4x4 m) than the Landsat TM raster (30x30 m), the two images can not be easily compared through image overlays in IDRISI (IDRISI user’s manual). The software requires that the overlay operation use raster images of the same resolution and extent (number of rows and columns). Therefore, the two rasters must be brought to the same resolution and size. This was accomplished using image expansion. Each pixel is subdivided into smaller pixels. For example, if expansion factor is 4, each pixel is subdivided into 4 x 4 = 16 pixels. Through image expansion, the IKONOS raster was converted to 2 m by 2 m pixels (expansion factor = 2, given the original pixel resolution of 4 m x 4 m). For the Landsat raster, the expansion factor was 15.

This statistical measure is an estimate of the difference between the observed accuracy and the probability of chance agreement of classes (Lillesand and Kiefer, 1994). The Kappa coefficient was estimated at 0.4979 for this cross tabulation. Barren land, water features, and dense coniferous forests have the best match between the two images. For other classes the match is poor.

It is very useful to visualize the errors between the two classification results. Two new images were created using the image algebra and overlay operations in IDRISI. The details are found in Project Diagram in the Appendix 3, and the results show images with four Boolean algebra operations: pixels classified as Fen in IKONOS but NOT in Landsat, pixels classified as Fen in Landsat but NOT in IKONOS, pixels classified as Fen in both IKONOS AND Landsat, and other pixels (the negation of the “AND” operation on IKONOS and Landsat images). The same logic follows for class of sparse coniferous forests, or any other class. Only class “Fen” and class “sparse coniferous forest” were analyzed here.