Abstract:
A technique for collecting and recording reference data which considers the spectral and spatial characteristics of Landsat data, the computer system being used, and the gradient nature of wildland vegetation was developed and described. Different analysis techniques for four critical factors affecting the accuracy of computer-aided analysis products were evaluated. Comparisons were made on the basis of accuracy evaluations of two methods of data/analyst interface, three methods of deriving training statistics, three methods of spectral class descriptions, and two levels of map category detail. The primary data set used was digital Landsat multispectral data for a study area around Fairbanks, Alaska. Reference data were developed from field work and photo-interpretation. The training methods compared were supervised, unsupervised, and modified clustering. The three spectral class description methods were: (1) labels derived from the training data; (2) the color display screen; and (3) from ground plot data. Community level cover types were compared with generalized map categories. The effect of post-classification stratification was evaluated. The reference data technique provides geographically located stands and cover types identifications with a flexible coding system that can be aggregated to correspond to the spectral data categories. No difference in classification accuracy was found for an experienced analyst using a printout oriented system such as EDITOR or a screen oriented system such as IDIMS. The modified cluster method of developing training statistics was more effective and efficient than supervised or unsupervised training methods. The use of ground plot data and subsequent stratification improved the descriptions of spectral classes. Generalized mapping categories were more accurate than detailed mapping categories. Knowledge of the ecologic, floristic, and spectral characteristics of the cover types in the study area is necessary to develop spectral class descriptions and stratification criteria.
