A Standardized Probability Comparison Approach for Evaluating and Combining Pixel-based Classification Procedures

In this paper, a standardized probability approach is presented to evaluate the pixel labeling confidence of each pixel and then combine the classification maps generated from different classification procedures for improving classification accuracy. This approach examines the posterior probability of the maximum-likelihood classifier or inverse-distance weight for the minimum-distance classifier for each pixel. It recommends that, for every classification, a standardized probability map should be outputted along with the classified map to show the pixel labeling confidence for all pixels. Tests based on different feature combinations and training strategies from Ikonos data show that the proposed approach was effective in improving the labeling confidence, as well as overall classification accuracy when classified maps from different classification procedures were combined. This standardized probability can be used to provide additional spatial information along with the traditional accuracy assessment.