LARS Tech Report Number
Dual polarized X-Band Synthetic Aperture Radar (SAR) data were obtained for a test site in South Carolina containing a variety of forest and agricultural cover types, as well as water and urban areas. The optically correlated images were digitized with a scanning microdensitometer using a 40 µm aperture. Digital registration of the two polarizations was more difficult than anticipated due to geometric variations in the imagery. However, a successful registration was obtained, and a "degraded" 30 m resolution data set was generated in addition to the original 15 m resolution data set.
Computer analysis indicated a statistically significant look-angle effect in the radiometric characteristics of the data. Three different classification algorithms were tested: (1) GML, or Gaussian Maximum Likelihood; (2) Per-Field; and (3) SECHO or Supervised Extraction and Classification of Homogeneous Objects. The GML is a per-point classifier, whereas the latter two are contextual classifiers, in that the classification decision is based on both the mean and the variance of the spectral response over an area. Evaluation of the classification results, based on test data for the seven major cover types present in the study site, indicated a significant improvement in accuracy for the contextual algorithms as compared to the GML per-point algorithm, but overall performance was only 65% for even the contextual algorithms. The effects of spatial resolution and polarization of the radar signal, as well as the classification algorithm, are discussed in this paper.
The results indicate the need for: (1) completely digital data processing of SAR data (as compared to optical correlation techniques for producing the SAR imagery); (2) evaluation of longer wavelength SAR data for differentiating forest and other cover types and condition classes; and (3) improvements in contextual algorithms and analysis techniques.
Date of this Version