Performance analysis of supervised image classification techniques for the classification of multispectral satellite imagery

Abstract

Remote Sensing is extensively used for crop mapping and management in current era. High resolution multispectral data of every part of earth is available at relatively low cost. Selection of appropriate decision rule and appropriate spectral bands is critical for obtaining accurate classification results. Need to find an accurate decision rule which is less time consuming and needs less resources leads to the performance analysis of different classification algorithms on the basis of their classification accuracy, time consumption, computational requirements, reliability etc. A SPOT 5 image acquired on 2009-07-03 from Barcelona city of Spain located at 41.3833° N, 2.1833°, and the classification algorithms i.e. Maximum Likelihood, Parallelepiped, and Mahalanobis Distance classifiers were used for the classification of the SPOT image. A spatial subset of the original imagery was created with resolution half of the original image. In this research, imagery was first atmospherically corrected using QUAC (Quick Atmospheric Correction). The spectral signature of different land classes were extracted using the spectral profile of each individual land class. The whole imagery was then classified using three Classifiers/Decision Rules i.e. Maximum Likelihood, Mahalanobis Distance and Parallelepiped Classifier. The post classification procedures i.e. clump and sieve were applied to the classified imagery to improve classification results. Maximum Likelihood Classifier outperforms other classifiers i.e. Mahalanobis Distance and Parallelepiped Classifier with Overall Accuracy (OAA) of 99.17 per cent. However these classifiers show good accuracy for classification of some classes of interest for instance the Mahalanobis Classifier outperforms the Maximum Likelihood Classifier in classifying water bodies. Results also show that the band selection is also critical in accurate classification of the imagery. The spectrally subsetted images (NIR band removed) of the same place showed very less classification accuracy than that of the original image