Land Cover Mapping Capability of Chaincluster, K-Means, and ISODATA techniques-A Case Study

摘要

Unsupervised classifiers have been consistently used in remote sensing studies and are particularly of a large advantage when the terrain information is not available. In this paper, we have analyzed three unsupervised classifiers in their ability to both identify and create meaningful LULC classes in the study area. All three classifiers have excellently extracted the spectrally and spatially dominant classes in the study area. However, a great difference is observed when it comes to extracting spectrally as well as spatially subservient classes. Particularly, the chaincluster technique failed to identify built-up class in the study area, even though it produced the same amount of overall classification accuracy as the k-means clustering technique. However, it should be noted that the overall Kappa value produced by the chaincluster technique is better than that produced by k-means, which indicates that the output thematic map produced by chaincluster is more efficient than that produced by the k-means technique. ISODATA clustering produced considerably better classification performance than both chaincluster and k-means techniques. The Kappa statistic for ISODATA is better by 0.14 compared to other two techniques and this is a great improvement in the overall thematic map quality. Hence, we conclude that the ISODATA classification technique is more suitable for identifying LULC classes than chaincluster and k-means techniques. Also, as per our second objective, the spectrally dominant and subservient classes were identified successfully. It would be interesting to test the ability of these classifiers applied to very high-resolution RS data such as hyperspectral imagery.