A comparison of maximum likelihood-based supervised classification strategies

Supervised land cover classification strategies for Thematic Mapper image data are tested in two southern California study sites. Effects of training field aggregation, prior probabilities, and nonnormality in spectral data are investigated. Findings are summarized using both PCC and Kappa statistics. Error matrices are presented in graphic form to enhance comparison among images and methods. Differences in overall map accuracy between tested strategies are generally significant, and certain strategies tend to produce consistent patterns in the error matrices of all image products. Results of these supervised classifications are also briefly compared to previously published results for unsupervised techniques using identical image data products.