Spectral similarity for evaluating classification performance of traditional classifiers

摘要

Image classification is the most extensively used technique for extracting useful information from the remotely sensed data. It is possible to extract highly accurate details like particulars of a land cover by effectively employing suitable classification techniques. Class separability is an important index in deciding classification accuracy. If a pair of classes is spectrally independent, the likelihood of producing misclassifications between them is narrow. The main aim of this study was to analyze the performance of three traditional classifiers over current technology heterogeneous multispectral LANDSAT-8 data. Normalized Euclidean distance was used as the principle index for defining class separability. The classification algorithms considered were: Maximum Likelihood, Minimum-distance-to-means and Mahalanobis distance. Accuracy assessment was carried out to statistically evaluate the performance of each classifier. It is observed from the study that each classifier has a threshold for class separability index below which a large number of misclassifications were produced during classification. A comparative analysis was also carried out to emphasize the differences in classification results.