ACCURACY ASSESSMENT OF REMOTELY SENSED DERIVED THEMATIC MAPS

Abstract

Accuracy assessment of remotely sensed derived thematic maps have lately become an integral part of remote sensing image classification. There are a number of measures for the evaluation of accuracy of thematic classifications both crisp and fuzzy that have been proposed in the remote sensing literature. There may be lot of variation in the results of classification by the use of different accuracy measures. However, the currently available commercial image processing packages incorporate only a few of these measures. An attempt has been made here to develop a software for assessing the accuracy of thematic maps. The package has been written in MATLAB script. In order to perform the classification in crisp and fuzzy modes, the algorithms for two classifiers namely, Maximum Likelihood and Fuzzy C-Mean have been included. All commonly used accuracy measures for crisp and fuzzy classification outputs have been considered. The software has been named as RSICAA and contains five basic modules: Display, Training Data, Classification, Testing Data and Accuracy Assessment Module. The performance of classifications has been evaluated using IRS IC LISS III data. A thorough comparison between various accuracy measures has been made. It has been observed that for the data set considered, the MLC and FCM in supervised mode is significantly better than that of FCM in unsupervised mode. Further, for accuracy assessment of crisp classifications, the Kappa and Tau coefficients appear appropriate, whereas for fuzzy classifications, measures of closeness may be considered better than others.