Abstract: The management of humanitarian operations in highly intense situations like migration movements happening at borders often lack current and sufficient information. Satellites do provide large-scale information fast. When dealing with a migration situation, satellite images now can give information about where refugees are before they arrive at a border, giving first responders urgently needed lead time for contingency and capacity planning. Dwelling Detection, a method conducted on satellite images of refugee camps, is able to count the dwellings in a camp. From that, the number of inhabitants in a camp can be derived for forecasting purposes. To count the dwellings, object detection machine learning methods can be used. In Wickert et al. [ASPAI' 2020, 1, 2020], a dwelling detection workflow using a Faster R-CNN is described. The workflow contains a newly developed annotation method, an inhabitant estimate for analyzed camps and a global confidence factor indicating the quality of the analysis of the image and the estimate of the inhabitants. In this actual extension of Wickert et. al. [ASPAI’ 2020, 1, 2020], lessons learned from multiple training and testing runs are documented, following a detailed analysis of those tests and validations in Wickert et. al. [ISPRS 2020, 2, 2020]. In this extended article we conclude that the workflow produces results that can be used in humanitarian operations. We further document our lessons learned in developing a dwelling detection workflow and we provide recommendations for training a dwelling detection classifier. We advise humanitarian operators to build a dwelling detection classifier following our recommendations and use satellite images in actual humanitarian operations. This approach can reduce stress for all people involved in a humanitarian (crisis) situation and lead to better decisions in intense migration situations.
Keywords: Artificial Intelligence (AI), Machine Learning (ML), Deep Learning, Convolutional Neural Network (CNN), Remote Sensing (RS), Dwelling Detection, Object Detection, Lessons Learned.
___________________________________________________________________