Thursday, September 30, 2021

Dealing with imbalanced data of forcibly displaced people’s dwellings

Suggested/supervised by: Yunya Gao, Stefan Lang

Due to human armed conflicts, human rights violations, persecution, or environmental degradation, numerous people are being forced to flee their homes globally. They are called refugees if they have crossed international borders, or internally displaced persons (IDPs) in the case they are on the run within their home countries. Refugee/IDP dwellings are commonly used as temporary shelters for these displaced people. Detailed information on refugee/IDP dwelling infrastructure, the population in need and their spatial distribution is important for planning humanitarian actions (Sprohnle, Fuchs, & Aravena Pelizari, 2017). However, during a crisis, such critical information is usually hard to access by fieldwork (Witmer, 2015). Therefore, detecting refugee/IDP camps through remote sensing (RS) techniques has attracted a lot of attention.

Very high spatial resolution optical (VHSRO) satellite imagery is considered as the major source of information to identify the number and types of structures (Ghorbanzadeh, Tiede, Wendt, Sudmanns, & Lang, 2021). Visual interpretation of VHSRO imagery is recognized to produce accurate refugee/IDP camp mapping results (Witmer, 2015). However, it is thought to be highly reliant on context knowledge, time-consuming and labour intensive (Witmer, 2015). Recently, deep learning (DL) methods have been proved to outperform many other automatic processing methods in many remote sensing domains. However, DL methods usually require a large amount of data and high computational cost (Zhu et al., 2017). In the past few decades, ZGIS has assisted Doctors without borders (Médecins Sans Frontières, MSF) in collecting a large amount of data of refugee/IDP camps together with VHSRO imagery for many countries. Therefore, DL methods have high potentials in detecting refugee/IDP dwellings. However, by looking through the collected data, there are usually multiple classes of dwellings. The amount of data for each class is imbalanced. After some initial experiments, it is found out that the accuracy of minority classes (e.g. Tukul) is usually much lower than majority classes (e.g. bright dwellings). It is still unknown how to deal with imbalanced data issues in refugee/IDP dwelling detection by using DL methods. Improving the accuracy of prediction results of minority classes may be valuable for planning humanitarian actions.

Figure 1. Three types of refugee/IDP dwellings represented from very high spatial resolution satellite imagery

Reference:

  • Ghorbanzadeh, O., Tiede, D., Wendt, L., Sudmanns, M., & Lang, S. (2021). Transferable instance segmentation of dwellings in a refugee camp - integrating CNN and OBIA. European Journal of Remote Sensing, 54(sup1), 127–140. https://doi.org/10.1080/22797254.2020.1759456
  • Sprohnle, K., Fuchs, E. M., & Aravena Pelizari, P. (2017). Object-Based Analysis and Fusion of Optical and SAR Satellite Data for Dwelling Detection in Refugee Camps. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 10(5), 1780–1791. https://doi.org/10.1109/JSTARS.2017.2664982
  • Witmer, F. D. W. (2015). Remote sensing of violent conflict: eyes from above. International Journal of Remote Sensing, 36(9), 2326–2352. https://doi.org/10.1080/01431161.2015.1035412
  • Zhu, X. X., Tuia, D., Mou, L., Xia, G. S., Zhang, L., Xu, F., & Fraundorfer, F. (2017). Deep learning in remote sensing: a review. ArXiv, (December). https://doi.org/10.1109/MGRS.2017.2762307

No comments: