Utilizing orthophotos for base map maintenance through building detection
No Thumbnail Available
URL
Journal Title
Journal ISSN
Volume Title
School of Engineering |
Master's thesis
Authors
Date
2024-12-19
Department
Major/Subject
Mcode
Degree programme
Master's Programme in Geoinformatics
Language
en
Pages
67
Series
Abstract
Managing the building inventory and keeping it up to date is a challenging task especially over large area. Orthophotos offer rich features to detect buildings such as textures, edges, corners and colours not to mention they can be enriched with other data like Lidar. Orthophotos are also georeferenced and orthorectified making precise mapping and comparison to existing map possible. The goal of the thesis was to find out whether missing buildings could be detected from orthophotos by comparing them to existing map. It was also explored how detected buildings could be determined to need updating. Study was conducted in the Helsinki area by having seven square kilometres of multichannel true orthophotos containing channels RGB + normalized digital surface model. Buildings were detected by utilizing Unet with ResNet34 as a backbone from Fastai and Pyrotch libraries. ImageNet weights were used for RGB channels. Training data was created with a script from multichannel true orthophotos that were manually labelled. The inference script was built to output predictions straight as georeferenced GeoTIFFs. A FME workspace was also built to read georeferenced pre-dictions and to compare them to existing base map. This study shows that missing buildings are possible to detect from orthophotos using deep learning methods and by modifying the output as georeferenced Geo-TIFFs. It was also found out that comparison to existing map is possible to do semi automatically by combining automatic and manual inspection. As a byproduct a pipeline to reproduce the study with different configurations was created.Description
Supervisor
Rönnholm, PetriThesis advisor
Gröhn, SimoKeywords
deep learning, orthophoto, building footprint, map updating, feature extraction, remote sensing