V.N. Alexandrov et al. (Eds.): ICCS 2006, Part III, LNCS 3993, pp. 506 – 513, 2006. © Springer-Verlag Berlin Heidelberg 2006 Detection of Tornados Using an Incremental Revised Support Vector Machine with Filters Hyung-Jin Son and Theodore B. Trafalis School of Industrial Engineering, The University of Oklahoma 202 W. Boyd, CEC 124, Norman, OK 73019, U.S.A. {son, ttrafalis}@ou.edu Abstract. Recently Support Vector Machines (SVMs) have played a leading role in pattern classification. SVMs are quite effective to classify static data in numerous applications. However, the use of SVMs in dynamically data driven application systems (DDDAS) is somewhat limited. This motivates the devel- opment of incremental approaches to handle DDDAS. In an incremental learn- ing approach, it is critical to keep a certain number of support vectors (SVs) without seriously sacrificing the generalization performance of SVMs. In this paper a novel incremental SVM method, called an incremental revised support vector machine with filters (IRSVMF) is proposed to resolve the above limita- tions. Computational experiments with tornado data show that this approach is quite effective to reduce the number of SVs and computing time and to increase the detection rate of tornados. 1 Introduction Support Vector Machines (SVMs) have played a leading role in pattern classification. Applying SVMs into the real world has motivated the development of incremental approaches to deal with huge data that are continuously coming to a learning system. Numerous publications point out that the standard SVMs cannot properly handle large-scale data sets and that the incremental approach is a remedy to overcome limi- tations of the standard SVMs [1, 2, 3]. In applying the SVMs approach in an incremental framework for classification problems, we will face several limitations as follows: First, support vectors (SVs) are accumulated as the incremental learning process is repeated. Therefore, it is important to control the number of SVs in SVMs. Second, SVMs waste most computing time for computation of kernel function values using less important data. If SVMs are used for training data from a particular classification problem such as an unbalanced classification problem in which there are many data in one class (less important) and few data in the other class (important), then use of computing time for kernel function evaluation among data points that are less impor- tant should be avoided to reduce the training time. The tornado detection problem is an application to be considered in this study. It can be characterized as follows: First, it is a two-class (tornado and non-tornado) clas- sification problem. Second, it is a problem with unbalanced data. The tornado class