Canadian Forest Service Publications
Inclusion of topographic variables in an unsupervised classification of satellite imagery. 2004. Wulder, M.A.; Franklin, S.E.; White, J.C.; Cranny, M.M.; Dechka, J.A. Canadian Journal of Remote Sensing 30(2): 137-149.
Issued by: Pacific Forestry Centre
Catalog ID: 24591
Unsupervised classification has emerged as a common method for mapping the land cover of large areas with satellite data. Typically, clusters generated by an unsupervised algorithm, such as the K-means, are merged and labelled using a combination of manual and automated methods. Topographic shadows found in areas of high relief, particularly in areas with low sun angles, increase the complexity of land cover classification, as a single land cover class may have very different spectral responses in shadowed and nonshadowed areas of the image. Methods to increase classification accuracies in areas with severe topographic shadows are required for Canadian large area land cover mapping projects. In a standard supervised classification, increases in land cover map classification accuracy have been obtained by including topographic attributes as inputs to the classification algorithm. In this study we investigate the potential of such data as a means to increase the accuracy of an unsupervised land cover classification in a high-relief area in central British Columbia, Canada. Separate datasets were used in cluster labelling and accuracy assessment. Four classification trials were completed: (i) using a standard approach without the addition of topographic attributes, (ii) using elevation data as an additional input to (i), (iii) prestratifying the image into shadow and nonshadow areas prior to clustering, and (iv) combining the methods used in (ii) and (iii). The latter provided the highest level of overall attribute accuracy at 80.1%, with a 95% confidence interval of 73.6%–88.6%. This is an improvement over the standard approach, which produced an overall attribute accuracy of 68.7%, with a 95% confidence interval of 59.8%–77.6%. We concluded that the prestratification of an image into areas of shadow and nonshadow prior to clustering in conjunction with the use of elevation data as an input to the clustering process is a practical method to increase classification accuracy in areas of high relief where topographic shadow is problematic.