University of Twente Student Theses

Login

Automatic road extraction from high-resolution remote sensing imagery using fully convolutional networks and transfer learning

Guthula, Venkanna Babu (2020) Automatic road extraction from high-resolution remote sensing imagery using fully convolutional networks and transfer learning.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Abstract:Road extraction from high-resolution remote sensing imagery is a fundamental task in remote sensing. Due to roads are very sophisticated features and complex background on the remote sensing imagery, the research has been continuously increasing on this topic. The various features such as trees, cars, buildings, and building shadows occlude the roads, and this causes reducing the accuracy due to continuity and gaps between the road segments. In this research work, we propose two novel methods to extract the road network from remote sensing imagery using ReuseNet that can be built with any fully convolutional network (FCN) architecture. Another method is extracting the road network using fully convolutional networks and transfer learning. Various experiments have been conducted. First, the road network extracted from two benchmark datasets using three state-of-the-art networks such as U-Net, SegNet, and ResUNet. And compared the results between each other and previous research work. And the transfer learning introduced to FCNs. In transfer learning, the learned feature from one dataset transferred to another network to learn quickly on the other dataset. And the ReuseNet method was used to improve the results of state-of-the-art networks. ReuseNet built with three state-of-the-art networks and extracted the road network from the two benchmark datasets and compared results with the results from state-of-the-art networks. All three networks are well performed on two datasets. Using transfer learning with FCNs reduces the computational time, and the accuracy almost closes to the model that trained with more epochs. And the ReuseNet improves the accuracy of the three state-of-the-art results.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:https://purl.utwente.nl/essays/98654
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page