Full text
| |
| Author(s): |
John Edgar Vargas Muñoz
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2019-10-04 |
| Examining board members: |
Alexandre Xavier Falcão;
Raul Feitosa;
Clarimar José Coelho;
Stanley Robson de Medeiros Oliveira;
Rubens Augusto Camargo Lamparelli
|
| Advisor: | Alexandre Xavier Falcão |
| Abstract | |
Geographical data are used in several applications, such as mapping, navigation, and urban planning. Particularly, mapping services are routinely used and require up-to-date geographical data. However, due to budget limitations, authoritative maps suffer from completeness and temporal inaccuracies. In this context, crowdsourcing projects, such as Volunteer Geographical Information (VGI) systems, have emerged as an alternative to obtain up-to-date geographical data. OpenStreetMap (OSM) is one of the largest VGI projects with millions of users (consumers and producers of information) around the world and the collected data in OSM are freely available. OSM is edited by volunteers with different annotation skills, which makes the annotation quality heterogeneous in different geographical regions. Despite these quality issues, OSM data have been extensively used in several applications (e.g., landuse mapping). On the other hand, it is crucial to improve the quality of the data in OSM such that applications that depend on accurate information become more effective (e.g., car routing). In this thesis, we review and propose methods based on machine learning to improve the quality of the data in OSM. We present automatic and interactive methods focused on improving OSM data for humanitarian purposes. The methods can correct the OSM annotations of building footprints in rural areas and can provide efficient annotation of coconut trees from aerial images. The former is helpful in the response to crises that affect vulnerable areas, while the later is useful for environmental monitoring and post-disaster assessment. Our methodology for automatic correction of the existing OSM annotations of rural buildings consists of three tasks: alignment correction, removal of incorrect annotations, and addition of missing building annotations. This methodology obtains better results than supervised semantic segmentation methods and, more importantly, it outputs vectorial footprints suitable for geographical data processing. Given that this automatic strategy could not attain accurate results in some regions, we propose an interactive approach which reduces the human efforts when correcting rural building annotations in OSM. This strategy drastically reduces the amount of data that the users need to analyze by automatically finding most of the existing annotation errors. The annotation of objects from aerial imagery is a time-consuming task, especially when the number of objects is high. Thus, we propose a methodology in which the annotation process is performed in a 2D space of projected image features. This method allows to efficiently annotate more objects than using traditional photointerpretation, collecting more effective labeled samples to train a classifier for object detection (AU) | |
| FAPESP's process: | 16/14760-5 - Interactive Annotation of Remote Sensing Images |
| Grantee: | John Edgar Vargas Muñoz |
| Support Opportunities: | Scholarships in Brazil - Doctorate |