LARS Tech Report Number
071482
Abstract
This study evaluated computer-aided analysis techniques applied to Thematic Mapper Simulator (TMS) data for the purpose of mapping forest cover types. Specifically, classification results obtained using a supervised set of training statistics and various combinations of three and four channels subsets of the seven available TMS channels are compared for three classification algorithms: L2, GML, and SECHO. In the analysis, the best three and four channel subsets were determined by minimum transformed divergence criteria. A Karhunen-Loeve or Principal Component linear transformation was applied to the 1979 TMS data set and supervised training statistics were generated for classifying the transformed data.
Classification results from applying the same three classification algorithms on the transformed data are compared to results from the untransformed data sets. Results from the untransformed TMS data show a higher performance using the four simulated Landsat channels (CH2:0.52 - 0.60µm; CH3: 0.63 - 0.69µm; CH4: 0.76 - 0.90µm; CH5: 1.00 -1.30µm) than from the best four channels selected by the minimum transformed divergence criteria. The contextual classifier known as SECHO (Supervised Extraction and Classification of Homogeneous Objects) performed significantly better than either of the two per-point classifiers for the untransformed data. Overall classification results of the K-L transformation increased for the L2 algorithm, but decreased for both the GML and SECHO algorithms.
Date of this Version
January 1982