Department of Geography and Environmental Studies
Jennifer M. Graham, 2005
ANALYSIS OF SURFACE COVER CLASSIFICATION USING IKONOS DIGITAL IMAGERY IN A COMPLEX HABITAT: POINTE DE L'EST, MAGDALEN ISLANDS
Remotely sensed digital images have become increasingly more accessible due to the launch of satellites such as IKONOS, which offers high spatial resolution digital images to the public. These digital images are often used to create surface cover maps from automated classification algorithms, however there are challenges associated with producing reliable and accurate surface cover maps with these methods. In this project an IKONOS digital image of Pointe de l'Est, a recurved spit that forms part of the Magdalen Islands, Quebec, was used to create seven surface cover maps using supervised and unsupervised classification algorithms. The resultant maps were analyzed to determine: i) map accuracy, ii) the relationship between supervised and unsupervised classifications, and iii) what classification approach would produce the most useful surface cover map. Two supervised classifications were produced with 23 and 19 classes, and overall accuracy of 54 and 60% respectively. Five unsupervised classification were produced with a minimum of 30 and maximum of 90 clusters. Superclasses were identified which contained up to 20 clusters, and were similar to the classification scheme used for the supervised classifications. However, superclasses tended to represent species assemblages as opposed to the individual cover types identified for the supervised classification scheme. An additional advantage of the unsupervised classifications was the ability to detect ‘mixed' or transitional classes. It was concluded that the most useful surface cover map would be produced by a hybrid approach. Some regions, in particular coniferous forest, the islands which serve as anchors for the spit and Spartina salt marsh, should be masked and classified with a supervised algorithm. The remaining regions, in particular wetlands, dunes, heath communities, and transitional zones should be masked and classified using an unsupervised algorithm.