Frequency-based contextual classification and gray-level vector reduction for land-use identification

Abstract

Attempts to map land use directly from higher spatial resolution satellite data with conventional computer classification techniques have proven to be ineffective. This is due to two facts. First, land use is a cultural concept. What we see on remote sensing imagery is only the physical evidence of land use as represented by combinations of land-cover types. Second, conventional classifiers employ only spectral information on a single-pixel basis. A large amount of spatial information is thus ignored.

In this research, a contextual classification method was developed to obtain land-use information. The number of gray-level vectors in multispectral space was reduced using a new data-reduction algorithm through rotating multispectral space into eigen space. As a result, the multispectral data were reduced to images of one feature dimension with the loss of relatively little information. Each gray-level vector-reduced image was then used in the frequency-based procedure to derive land-use information.

These land-use classification procedures were tested using SPOT HRV data obtained over part of the rural-urban fringe of Metropolitan Toronto, Canada. Best overall classification accuracies (measured by the Kappa coefficients) obtained using the three procedures were 0.616 when a classification scheme with 14 land-use classes was used. These accuracies are significantly better than an accuracy of 0.462 which was obtained using the maximum-likelihood classification method. The contextual classifier developed proved to be very efficient in terms of computation.