Classification, change-detection and accuracy assessment: Toward fuller automation.

Nancy E. Podger - 2004
[UMI Proquest Full Citation]

This research aims to automate methods for conducting change detection studies using remotely sensed images. Five major objectives were tested on two study sites, one encompassing Madison, Wisconsin, and the other Fort Hood, Texas. (Objective 1) Enhance accuracy assessments by estimating standard errors using bootstrap analysis. Bootstrap estimates of the standard errors were found to be comparable to parametric statistical estimates. Also, results show that bootstrapping can be used to evaluate the consistency of a classification process. (Objective 2) Automate the guided clustering classifier. This research shows that the guided clustering classification process can be automated while maintaining highly accurate results. Three different evaluation methods were used. (Evaluation 1) Appraised the consistency of 25 classifications produced from the automated system. The classifications differed from one another by only two to four percent. (Evaluation 2) Compared accuracies produced by the automated system to classification accuracies generated following a manual guided clustering protocol. Results: The automated system produced higher overall accuracies in 50 percent of the tests and was comparable for all but one of the remaining tests. (Evaluation 3) Assessed the time and effort required to produce accurate classifications. Results: The automated system produced classifications in less time and with less effort than the manual ‘protocol’ method. (Objective 3) Built a flexible, interactive software tool to aid in producing binary change masks. (Objective 4) Reduced by automation the amount of training data needed to classify the second image of a two-time-period change detection project. Locations of the training sites in ‘unchanged’ areas employed to classify the first image were used to identify sites where spectral information was automatically extracted from the second image. Results: The automatically generated training data produces classification accuracies similar to accuracies from a classification where training data were manually collected. (Objective 5) Decrease the effort needed for post-classification change analysis. Classification accuracy metrics, produced from a hybrid change detection analysis for the second time period, were found to be comparable to results generated from a conventional classification of the same image. Further research showed that pixel-by-pixel comparisons between classifications produced conflicting results.