Spatial transferability of deep learning models for dwelling extraction in refugee camps and IDPs

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

Globally there is a profound number of the global population displaced from their homes because of natural and anthropogenic reasons. These forcibly displaced segment of the global population resides either in IDP sites or refugee camps if they cross international borders. Dwelling extraction and camp monitoring for a specific single camp can be efficiently done by using deep learning models and high-resolution satellite images and annotated labels to train the model. The biggest problem is when new IDPs/refugee camps are happening and rapid information retrieval is needed for emergency response for humanitarian response. The proposed work intends to leverage the learned skills of the model trained in IDP/refugee camps to extract features in completely unseen images taken from other new IDP/refugee camps. One of the constraints which limit this learning process is that dwelling characteristics in different refugee camps would be completely different which is governed by the medium of the material used to construct the houses, its contrast with the surrounding information, sizes of houses, and spacing of houses (See Fig. 1). A model trained in one site is said to be universally capable of extracting features if and only if it is spatially transferable and able to work very well in refugee camps that are located in other geographic locations. Therefore, the objective of the proposed work will be firstly assessing the spatial transferability of the instance segmentation model and designing implementation strategies that can improve the spatial transferability of the model.

Figure 1: Dwelling type and structures in three different refugee camps