| Grant number: | 16/14760-5 |
| Support type: | Scholarships in Brazil - Doctorate |
| Effective date (Start): | November 01, 2016 |
| Effective date (End): | June 30, 2019 |
| Field of knowledge: | Physical Sciences and Mathematics - Computer Science |
| Principal Investigator: | Alexandre Xavier Falcão |
| Grantee: | John Edgar Vargas Muñoz |
| Home Institution: | Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil |
| Associated scholarship(s): | 17/10086-0 - Interactive rural building detection and delineation using remote sensing images, BE.EP.DR |
Abstract Several remote sensing image classification methods have been proposed to generate thematic maps of land cover and urban areas. However, photo-interpreters still manually annotate remote sensing images - an exhaustive and error-prone task, especially when it relies on some artificial color image composition with hardly distinguishable classes. In land cover mappings, the problem not rare makes necessary to visit the geographical location for correct image annotation. In this context, the quality of the selected image features and training samples is crucial to minimize the user effort and time in image annotation. The use of regions as samples and active learning techniques can mitigate the problem by reducing the numbers of training samples and iterations for the expert's label supervision. However, the development of active learning methods often relies on the simulation of the expert's actions using pre-annotated data sets. Such practice ignores possible annotation errors and disregard the user as part of the learning loop. Moreover, the methods usually ignore that non-interest regions in the image must also be treated. We propose to investigate active learning methods that do not suffer from those shortcomings in order to develop a more realistic framework for remote sensing image annotation. The proposal relies on effective prior superpixel segmentation, feature extraction, and data organization in order to facilitate the active learning loop, where the apprentice classifier must select the most effective training samples for its learning process. For development and validation, we will use natural color images to facilitate label supervision with the expert being part of the active learning loop. The simulation of the expert's actions on color composition images, pre-annotated by specialists, will also be considered for a more extensive evaluation. | |