Multitemporal Sentinel-1A data for urban land cover mapping using deep learning: Preliminary results
2016 (English)In: European Space Agency, (Special Publication) ESA SP, European Space Agency, 2016Conference paper (Refereed)
The objective of this research is to evaluate multitemporal Sentinel-1A SAR data for urban land cover mapping using a pixel-based Deep Belief Network (DBN) and an object-based post-processing. Multitemporal Sentinel-1A SAR in both ascending and descending orbits were acquired in Stockholm during the 2015 vegetation season. The images were first terrain corrected, co-registered, speckle filtered and scaled to 8 bit. Then the images were segmented using KTH-SEG, an edgeaware region growing and merging algorithm. For classification, a pixel-based deep belief network (DBN) was used. Then classification result was post-processed using object-based majority voting. For comparison, the same dataset was classified using an object-based support vector machine (SVM). The preliminary results show that the hybrid deep learning classification scheme produced comparable results as object-based SVM while yielded higher accuracies for builtup classes.
Place, publisher, year, edition, pages
European Space Agency, 2016.
Deep belief network, Deep learning, Segmentation, Sentinel-1A SAR, Urban land cover classification, Image segmentation, Mapping, Orbits, Pixels, Classification results, Classification scheme, Deep belief network (DBN), Deep belief networks, Sentinel-1, Urban land cover mappings, Support vector machines
Electrical Engineering, Electronic Engineering, Information Engineering
IdentifiersURN: urn:nbn:se:kth:diva-195000ScopusID: 2-s2.0-84988468544ISBN: 9789292213053OAI: oai:DiVA.org:kth-195000DiVA: diva2:1044148
Living Planet Symposium 2016, 9 May 2016 through 13 May 2016
QC 201611022016-11-022016-11-012016-11-02Bibliographically approved