Analysis on the influence of multiclass strategies and kernel functions on support vector machine for remote sensing image classification

• Grant number: 16/02875-2
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: June 01, 2016
• End date: May 31, 2017
• Field of knowledge: Physical Sciences and Mathematics - Geosciences
• Principal Investigator: Rogério Galante Negri
• Grantee: Luccas Zambon Maselli
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Estadual Paulista (UNESP). Campus de São José dos Campos. São José dos Campos , SP, Brazil

Abstract

Among different techniques proposed in literature for remote sensing classification, a Support Vector Machine (SVM) have been employed. Usually, to SVM training are previously selected a single multiclass strategy and Kernel Function that supposedly are suitable to deal with the classification problem. The multiclass strategies are employed when the classification involves more than two classes. In this case the original problem is decomposed into binary sub-problems. The Kernel functions are employed to improve the classification accuracy. The choose of a specific multiclass strategy and a kernel function play directly influence on the classification accuracy result. One possibility, in addition to adopt a global multiclass strategy and a kernel function, is to set a kernel function for each binary sub-problem defined by the employed multiclass strategy. It is believed that this way provides more accurate results compared with the usual approach. Starting from this motivation, in a first stage, this project aims to study and compare different multiclass strategies and kernel functions. Furthermore, it is proposed a different architecture to train the SVM method, which will allow select a kernel function for each binary sub-problem defined by the multiclass strategy. This study will make the application of theoretical concepts studied and developed for the classification of land cover and land use in an Amazonian region. (AU)