A simple cleaning procedure for improvement of training sitestatistics

摘要

For improving the effectiveness of supervised training, a cleaningprocedure which operates by selectively dropping training site pixelsbased on the Mahalanobis distance and class probability has beenproposed. The method is iterative and takes into account the spectraloverlap in all image hands. The procedure results in greaterclassification accuracy with a narrower confidence interval. TheBhattacharrya distance measure of class separability improved from anaverage value of 1.9373 to 1.9797 with a maximum change for a class pairfrom 1.2671 to 1.9052. The overall classification accuracy increasedfrom 94.74±0.64 to 39.63±0.19