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

This paper proposes a novel domain adaptation algorithm to handle the\nchallenges posed by the satellite and aerial imagery, and demonstrates its\neffectiveness on the built-up region segmentation problem. Built-up area\nestimation is an important component in understanding the human impact on the\nenvironment, the effect of public policy, and general urban population\nanalysis. The diverse nature of aerial and satellite imagery and lack of\nlabeled data covering this diversity makes machine learning algorithms\ndifficult to generalize for such tasks, especially across multiple domains. On\nthe other hand, due to the lack of strong spatial context and structure, in\ncomparison to the ground imagery, the application of existing unsupervised\ndomain adaptation methods results in the sub-optimal adaptation. We thoroughly\nstudy the limitations of existing domain adaptation methods and propose a\nweakly-supervised adaptation strategy where we assume image-level labels are\navailable for the target domain. More specifically, we design a built-up area\nsegmentation network (as encoder-decoder), with an image classification head\nadded to guide the adaptation. The devised system is able to address the\nproblem of visual differences in multiple satellite and aerial imagery\ndatasets, ranging from high resolution (HR) to very high resolution (VHR). A\nrealistic and challenging HR dataset is created by hand-tagging the 73.4 sq-km\nof Rwanda, capturing a variety of build-up structures over different terrain.\nThe developed dataset is spatially rich compared to existing datasets and\ncovers diverse built-up scenarios including built-up areas in forests and\ndeserts, mud houses, tin, and colored rooftops. Extensive experiments are\nperformed by adapting from the single-source domain, to segment out the target\ndomain. We achieve high gains ranging 11.6%-52% in IoU over the existing\nstate-of-the-art methods.\n