Land cover classification models using shuttle imaging radar (sir-c) data: A case study in new hampshire, u.s.a.

Spaceborne synthetic aperture radar (SAR) systems have the ability to provide high resolution information on land cover characteristics under adverse conditions such as darkness or cloud cover. The use of multiple frequencies and multiple polarizations yields better classification accuracies. The results of various land cover classification algorithms using Shuttle Imaging Radar (SIR-C) SAR data as applied to a site in Suncook, New Hampshire, are described in this paper. Three classification models were developed and tested: minimum distance classification, maximum a posteriori probability classification, and neural network classification. Using the available ground truth information, land cover was classified into five distinct regions: water, swamp, sand, trees, and grass. All three methods were applied to the same site and results compared. The maximum a posteriori probability approach yielded the highest overall classification accuracy on a pixel-by-pixel basis. Although the minimum distance approach was simpler than the maximum a posteriori approach, its performance was not as good as the latter since it did not use the covariance information between the data channels. The neural network approach performed well and its results were comparable to the maximum a posteriori approach when the variance of the data was small; however, its performance degraded rapidly when the variance of the data was high.