Peculiarities of SVM-Based Classification of BPG Compressed Three-Channel Images

Remote sensing systems provide a great amount of useful data for various applications. To transfer these data via communication lines and/or to store them, compression is applied. Lossy compression techniques are used more often since a variable and quite large compression ratio can be attained. Meanwhile, introduced distortions usually result in reduction of probability of correct classification. Then, it is desired to establish connection between compressed image quality and classification characteristics. The latter also depend on a used coder and an applied classifier. In this paper, we study the influence of compression carried out by better portable graphics (BPG) coder applied to three-channel images where a trained support vector machine (SVM) classifier is employed. Several quality metrics describing image quality and parameters that deal with classification accuracy are used. We show that reduction of classification accuracy is almost negligible for small values of parameter Q that controls introduced losses and it starts to quickly increase for Q > 17. Dependences of classification accuracy on Q are individual and depend on image properties. Some dependencies have fluctuating character. The recommendations on Q setting are given.