eduzhai > Applied Sciences > Engineering >

Weakly Supervised Domain Adaptation for Built-up Region Segmentation in Aerial and Satellite Imagery

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 29 pages

Abstract: This paper proposes a novel domain adaptation algorithm to handle thechallenges posed by the satellite and aerial imagery, and demonstrates itseffectiveness on the built-up region segmentation problem. Built-up areaestimation is an important component in understanding the human impact on theenvironment, the effect of public policy, and general urban populationanalysis. The diverse nature of aerial and satellite imagery and lack oflabeled data covering this diversity makes machine learning algorithmsdifficult to generalize for such tasks, especially across multiple domains. Onthe other hand, due to the lack of strong spatial context and structure, incomparison to the ground imagery, the application of existing unsuperviseddomain adaptation methods results in the sub-optimal adaptation. We thoroughlystudy the limitations of existing domain adaptation methods and propose aweakly-supervised adaptation strategy where we assume image-level labels areavailable for the target domain. More specifically, we design a built-up areasegmentation network (as encoder-decoder), with an image classification headadded to guide the adaptation. The devised system is able to address theproblem of visual differences in multiple satellite and aerial imagerydatasets, ranging from high resolution (HR) to very high resolution (VHR). Arealistic and challenging HR dataset is created by hand-tagging the 73.4 sq-kmof Rwanda, capturing a variety of build-up structures over different terrain.The developed dataset is spatially rich compared to existing datasets andcovers diverse built-up scenarios including built-up areas in forests anddeserts, mud houses, tin, and colored rooftops. Extensive experiments areperformed by adapting from the single-source domain, to segment out the targetdomain. We achieve high gains ranging 11.6 -52 in IoU over the existingstate-of-the-art methods.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×