Evaluation Of Multiple Classifiers System With Dimension Reduction Method Of Hyperspectral Image Classification For Land Cover Mapping

Supervised classification of hyperspectral image is considered to be the method of choice for improved land use/cover mapping. However, there is no single classifier which offers acceptable results across various study sites, sensors and resolutions. Further, the use of dimensionality reduction methods further increase the permutations and combinations required for selecting the appropriate classification algorithms. This has led to the identification of suitable combination of classifier and dimensionality reduction method in various hyperspectral image classifications. Multiple classifier system (MCS), a modern pattern recognition approach, has the ability to combine relative merits of various algorithms in a single classification task and hence increasing accuracy of image classification .This study has assessed the possibility of increasing the quality of hyperspectral image classification using the MCS by deploying dimensionality reduction for land cover classification. Multi-source airborne hyperspectral images acquired over four different sites covering a range of land cover categories have been classified by a MCS and compared against the classification results obtained from support vector machines (SVM). . It has been observed that the MCS offers superior classification results when there are multiple dimensionality reduction methods included in the MCS architecture. Apart from offering acceptable classification results, the MCS indicates about 1% increase in the overall accuracy and up to 1.3% increase in per class producer and user accuracy in comparison to individual classifiers accuracy. Overall accuracy estimates of the MCS are comparable SVM. However, much instability in the classification has been observed when the SVM is also included in the MCS architecture as base classifier.