The Effects of Polygon Boundary Pixels on Image Classification Accuracy

摘要

The purpose of this study is to analyze the effects that pixels, located at polygon boundaries, have on classification accuracy. Pixels found along the borders of polygons usually contain mixed spectral information, and can be detrimental to classification accuracy. Discriminant analysis was used to predict land cover classes, found in Canada's National Forest Inventory, from a Landsat TM image. The discriminant criteria were derived on a test area of the image, using buffered and non-buffered polygons as training data, and applied to a validate area of the image. Buffering the polygons had no overall positive or negative effect on classification accuracy. There are non-trivial effects for specific cover types, especially the water categories, but the classification accuracy for most categories changed by less than 10% due to buffering. Overall accuracy is quite low as well, usually less than 50%, which suggests that discriminant analysis may not be suited for predicting National Forest Inventory land cover classes from Landsat TM images.