PROMOTING ACCESS TO AFRICAN RESEARCH

Global Journal of Environmental Sciences

Log in or Register to get access to full text downloads.

Remember me or Register



Supervised remote sensing image classification: An example of a one-against-one multi-class polynomial kernel based support vector machine

O Okwuashi, E Eyo, A Eyoh

Abstract


Software like ILWIS and GRASS GIS can be employed for remote sensing image processing and geographic information systems applications. The modules of the aforementioned image processing software are based on conventional multi-class classifiers/algorithms such as maximum
likelihood classifier. These conventional multi-class classifiers/algorithms are usually written in programming languages such as C, C++, and python. The objective of this research is to experiment the use of a binary classifier/algorithm for multi-class remote sensing task, implemented in MATLAB. MATLAB is a programming language just like C, C++, and python. In this research, the support vector machine binary classifier/algorithm based on a one-against-one approach implemented in MATLAB is
applied to remote sensing multi-class problem. Both simulated and empirical satellite remote sensing data are used to train and test a one-against-one support vector machine classifier. For the purpose of
validating the experiment, the resulting classified satellite image is compared with the ground truth data. The polynomial kernel function is used for the modelling. In the simulated application, 25 pixels are used
for the experiment, out of which 6 pixels are used for training while 19 pixels are used for testing. Out of the 19 tested pixels 18 pixels are correctly classified while only 1 pixel is left unclassified. In the empirical application, 256 and 7182 pixels are unclassified and misclassified respectively out of a total of 62500 pixels; and the computed overall accuracy of the experiment is 88.1%. The satisfactory result of the experiment indicates substantial agreement between the classification result and the reference data.



AJOL African Journals Online