Land Cover Classification of Urban Areas: A Comparison of Object-Based and Pixel-Based Approaches
Author
Trock, Casper Frederik Stub
Term
4. term
Publication year
2018
Submitted on
2018-04-24
Pages
68
Abstract
This thesis compares object-based and pixel-based approaches for classifying urban land cover from high-resolution satellite imagery. Using Sentinel-2 scenes from 2015 and 2017 over Copenhagen, land cover maps were produced with support vector machines and random forests, first at the pixel level and then after segmenting the images into objects. Multiple segmentation parameter settings were tested to identify configurations that maximize classification accuracy. The resulting maps were used to analyze land cover change in TerrSet, and a digital surface model derived from LiDAR was added in a smaller test area to evaluate whether height information improves performance. The object-based approach achieved notably higher overall accuracy, with marked gains for built-up and grassland classes, while the LiDAR-derived surface model yielded localized improvements. Change detection indicated a general increase in grassland between 2015 and 2017, with net declines in other classes.
Specialet sammenligner objektbaserede og pixelbaserede tilgange til klassifikation af arealdække i byområder ud fra højopløselige satellitbilleder. Med Sentinel-2-scener fra 2015 og 2017 over København fremstilles arealdækkekort ved hjælp af support vector machine og random forest, først på pixelniveau og dernæst efter segmentering af billederne i objekter. Flere segmenteringsparametre afprøves for at finde de indstillinger, der giver højest nøjagtighed. Kortene anvendes efterfølgende til at analysere ændringer i TerrSet, og en digital overflademodel fra LIDAR inddrages i et mindre område for at vurdere, om højdeinformation kan forbedre resultaterne. Den objektbaserede metode gav en tydelig forbedring af den samlede nøjagtighed, især for klasserne bebygget og græs, mens LIDAR-baserede højdedata førte til forbedringer i udvalgte områder. Ændringsanalysen peger på en generel stigning i græsarealer mellem 2015 og 2017 og et nettofald i de øvrige klasser.
[This apstract has been generated with the help of AI directly from the project full text]