A Comparative Study of Supervised Classification Techniques for Multi-Spectral Images

Document Type : Original Article


Egyptian Armed Forces.


Classification of satellite images is an important key for ground features extraction and thematic maps production. Satellite images with multi-spectral bands provide rich data which is useful for features extraction and description. Many supervised classification methods have been developed for classifying the multispectral images. Each method has its own advantages and disadvantages (limitations). In this paper the performance of four of the common used supervised classification techniques is compared. The techniques considered here are: Parallelepiped (PP), Minimum Distance (MD), Mahalanobies (MA), and Maximum Likelihood (ML). They are applied on a set of multispectral images acquired by Worldview-2 satellite. The classification results accuracy are analyzed and evaluated The research work flow is processed by using ENVI. The developed maps are then visually compared with each other and accuracy assessments utilizing ground-truths. The assessment of classification results is represented in confusion matrix format and determination of Kappa coefficients. The preliminary results show that Maximum Likelihood (ML) gives accurate classification result for the area of study with overall accuracy 91.5741% and it is evaluated by Kappa coefficient which is 0.8846: