Learning from grey data: implication of input quality on deep learning-based dwelling extraction in IDPs/Refugee camps

Potential supervisory team: Stefan Lang, Dirk Tiede, Lorenz Wendt, Getachew Gella

Deep learning-based dwelling extraction in internally displaced population sites (IDPs) and refugee camps, satellite images could be obtained with different bit depth, processing labels (calibration, atmospheric correction, and further enhancement). With the same token, labeled/annotated data may also be obtained from crowd sources or in-house generated databases with a different label of quality where there might be class mislabelling, imprecise dwelling boundaries (Fig 1), completely unlabelled objects, and others. Therefore, the main theme of the proposed topic will be investigating the implication of investing intensive reprocessing phases on the image side or simply using a co-registered image on the performance of dwelling extraction. Undertaking intensive image pre-processing steps and the generation of the very precise respective object labels for model training and testing demands a profound amount of time. Especially when it comes to operational settings in emergency response for humanitarian response, it could be a pressing challenge for rapid mapping and information retrieval. Therefore, the proposed study will focus on two issues. Firstly, a thorough investigation will be done on trade-offs of using a clean but small amount of labeled sample data and a large number of grey samples with less precise annotation quality for dwelling extraction using deep learning models. Second, a substantial effort would be invested in the creation of robust deep learning strategies that could leverage from large grey data in a learning phase

Figure 1: Satellite image of dwelling features(A) and image with annotated labels with shifted object boundary and two mislabeled objects