Integration of Multiclass Strategies and Different Kernel Functions into Support Vector Machines for Remote Sensing Image Classification
Although several image classification methods have been proposed in literature, Support Vector Machine (SVM) is widely used in Remote Sensing applications. In addition to its robust mathematical formulation, the possibility of using different kernel functions and multiclass strategies highlights the...
| Authors: | , |
|---|---|
| Format: | article |
| Status: | Published version |
| Publication Date: | 2019 |
| Country: | Brasil |
| Institution: | Universidade Estadual Paulista (UNESP) |
| Repository: | Repositório Institucional da UNESP |
| Language: | Portuguese |
| OAI Identifier: | oai:repositorio.unesp.br:11449/246929 |
| Online Access: | http://dx.doi.org/10.14393/rbcv71n1-47208 http://hdl.handle.net/11449/246929 |
| Access Level: | Open access |
| Keyword: | Image Classification Kernel Functions Multiclass strategy Support Vector Machines |
| Summary: | Although several image classification methods have been proposed in literature, Support Vector Machine (SVM) is widely used in Remote Sensing applications. In addition to its robust mathematical formulation, the possibility of using different kernel functions and multiclass strategies highlights the attractiveness of this method. While kernel functions make possible to enhance the classification performance face to non-linearly separable data, multiclass strategies extend the original formulation of SVM in order to cope with problems involving more than two classes. However, it worth mention that particular choice involving a kernel function and a multiclass strategy implies directly on the classification performance. Furthermore, the best choice may be not a simple task. In order to reduce the freedom degree that arises from different possible combinations between kernel function and multiclass strategy, two architectures to training SVM are proposed. Three case studies involving land use and land cover classification with images acquired by different sensors are carried in order to verify the potential of presented architectures in comparison to usual approaches. |
|---|