A SAMPLING DESIGN SOFTWARE FOR IMAGE CLASSIFICATION ACCURACY

Abstract

Sampling is necessary for various jobs in Civil Engineering, wherever a large amount of data is analysed. Since collection of whole data set is a time consuming and costly affair, sample of representative data set arc generally collected. For a proper representative sample, it is necessary to understand the sampling procedure so that the decision can be made with a certain degree of confidence. There are various sampling schemes, each having its own advantages and disadvantages. Depending upon the problem in hand, a particular sampling scheme may be used mainly from accuracy and economic considerations. For a particular sampling scheme the sample size may be computed in several ways. With a proper sample size for a given sampling scheme, accurate estimate may be obtained with a certain tolerable amount of error. Despite the utility of sampling design, there seems to be lack of suitable software for a proper sampling design procedure. In the present work a software SDICAA has been developed for Sampling Design for Image Classification Accuracy Assessment. Though, the software has been written for its use in remote sensing studies (an important Civil Engineering Application) it can be utilised for other applications also. It is user-friendly software written in C language and can run on a PC with minimum 8 MB RAM and VGA card. The software accommodates four major aspects of sampling design for image classification accuracy assessment. ( ) Sampling schemes; five schemes namely simple random sampling, stratified random sampling, systematic sampling, stratified systematic unaligned sampling and cluster sampling are included. (ii) Sample size; both user defined and formulae based. (iii) Plotting of samples. (iv) Kappa coefficient as accuracy measure for different sampling schemes. In order to test the performance of the software, some experiments were planned. The accuracy of a classified image was evaluated using testing data collected, in four sampling schemes with variation in sample sizes. It was concluded that for the data set selected the variation in sampling scheme and sample sizes had insignificant effect on the accuracy of classification. However, the sample collected in systematic sampling produced the highest accuracy. Moreover, the sample size computed on the basis of proper formulation produced higher accuracies than those obtained form arbitrarily chosen sample sizes. Further, the same levels of accuracies were obtained with lower sample sizes than those achieved with higher sample sizes. This indicates that it is not the question of bigger the better.