Generating land-cover maps from remotely sensed data: manual vectorization versus object-oriented automation

Manual vectorization of multispectral images is a widely used method for making land-use or land-cover maps. Although it is usually considered relatively accurate it is very time consuming, which has prompted the use in recent years of various semiautomatic methods for classifying remotely sensed images. One of the most promising of the latter is object-oriented image analysis based upon image segmentation, but the accuracy of its results, as well as its time demands, are disputed. Accordingly, this paper compared manual vectorization with object-oriented classification to reveal the strong and weak points of each. Two qualitatively different datasets were classified using both methods; time costs were monitored and accuracy levels were compared. It was found that manual vectorization achieved better overall accuracy (up to 93% versus 84%), but the semiautomatic method was usually more accurate when classifying some specific features such as roads, built-up areas, broadleaf trees and coniferous trees. The verdict regarding time-efficiency was less clear cut. The best method depends upon the spatial and spectral resolution of the data being processed.