Supervised and Unsupervised Classification Performance Satellite Images

One of the main purposes of remote sensing satellite images is to interpret the observed data and classify features. Satellite image classification plays a major role in the extraction and interpretation of valuable information from massive satellite images.The primary purpose of this research is to classify multi-spectral Thematic Mapper satellite images using supervised classification. Unsupervised (RGB color model) and supervised classification (maximum likelihood method) is adopted to achieve the classification purpose. The classification accuracy depends upon the selection accuracy of the training area.PCA Transform is adopted and applied on the original bands to create the principal component images.  The first three principle component image contains most the information in all the original bands. For this purpose the first three principle component images are chosen as RGB images to create a colored image This colored image has been employed for determining and selecting the training areas which are very important for supervised classification after applying histogram equalization enhancement method on the colored image to make the selection of the training area more clarity and accuracy. After that the selection of the training areas is ready as an input for the supervised classification.Our results showed that the image with higher variance value doesn't represent a prerequisite in image clarity. The variance with mean value may reflect the quality of the image. The variance and means look like the torque in physics. So that we can see that the image with low variance and mean value near the middle of the dynamic range value has more quality than the image with high variance and mean value near the edge of the minimum or maximum value. Improve the selection of the training area (region of interest) visually plays an essential role in increasing the accuracy of supervised classification and this reflects on the calculation of the area for each class in the scene. In the present work, the overall accuracy increased from 68.5% to 98.9%  after adopting the histogram equalization technique as an enhancement technique in improving the selecting the training area visually.Although a high classification accuracy with principal component image, but still the classification process with original bands is better because its values represent the real spectral reflectance. The result of classification accuracy with the original bands is enhanced from 68.5% to 97.08% and became very comparable with the result of classification accuracy with the principal component images 98.86% when the selection of the training area (region of interest) has been improved visually. In this thesis, the ENVI (Environment for Visualizing Images) software version 4.5 has been used to achieve the aim of this study.