Road detection via deep learning: unsupervised domain adaptation for urban and suburban environment.

• Grant number: 24/02040-4
• Support Opportunities: Scholarships in Brazil - Master
• Start date: August 01, 2024
• End date: February 28, 2025
• Field of knowledge: Physical Sciences and Mathematics - Geosciences - Geodesy
• Principal Investigator: Aluir Porfírio Dal Poz
• Grantee: Gustavo Rota Collegio
• Host Institution: Faculdade de Ciências e Tecnologia (FCT). Universidade Estadual Paulista (UNESP). Campus de Presidente Prudente. Presidente Prudente , SP, Brazil

Abstract

The road network is one of the main ways of transport in the world, which provides many applications, such as urban planning, economy (logistics), autonomous navigation etc. Through aerial images, the road detection task is one of the most challenging and recurrent research topics in the Remote Sensing and Photogrammetry (RS&P) community, with the Convolutional Neural Networks (deep learning), the state-of-art. However, due to contextual aspects of roads and the necessity of a large labeled dataset, the model inference may still be limited. For example, roads have different scenarios, like pavemented or no pavemented roads, urban or rural areas etc., implying the characteristics and proprieties that define a domain for training can be distinct from a domain designed for testing - source domain and target domain. A manner of mitigating this shift between domains, which is known as domain shift in the literature, would be to label the objects of the target domain and train the model again. However, this elucidates the limitation that is habitual in RS&P, i.e. the availability of raw data, without labels. To reach these aforementioned challenges, this project proposes the use of unsupervised domain adaptation adapted for U-Net architecture, aiming to detect roads in urban and suburban environments. Specifically, the adaptation needs object labels just in the source domain, to infer the model in the target domain without labels necessity. Few studies have been carried out using this approach and they still suggest several challenges. Therefore, this becomes an opportunity for research in order to obtain higher accuracy of road network through deep learning.