Each year, 1.3 billion dollars in humanitarian aid is lost in refugee camps due to poor supply management. As a result, an estimated 1.88 million children, women and men each year do not receive the essentials to keep them safe and in good health. We use satellite images to automatically estimate the number of expected refugees for the Camp Forecast tool. We use machine learning to analyze these satellite images of refugee camps to improve the population estimations for more effective use of resources.
Collaboration with Elva
In collaboration with Elva, we developed the Camp Forecast tool for IOM to improve forecasting requirements and thus save time, reduce costs and possibly even save lives. IOM has decades of leading humanitarian experience and ELVA has experience collecting data in 20 conflict-affected countries around the world.
Currently, very long surveys are completed every month by camp managers. We started to create a tool that could address the most common deficiencies in these studies. In the tool, only the number of refugees in a camp and the existing stock need to be entered. The tool then calculates the required stocks and budget for the current situation, but also for several months in advance. However, we still need the number of refugees in a camp as input. This number can be taken from the monthly surveys. However, these surveys appear online with a delay of up to a month. So the question was: how do you get this information faster?
In order to estimate the number of people in a refugee camp closer to real time, satellite images of these camps can add useful information. This has led to two questions:
Object Detection for recognizing tents
To answer these questions, the following steps are performed:
The current state-of-the-art Object Detection models are based on Neural Networks. An important part of training Neural Networks is the preprocessing of the input. In this case we use satellite images from Google Earth as input. A lot of performance gains can be achieved by carefully examining, cleaning and transforming the input data. Subsequently, we investigated with how many 'epochs' the model had to be trained. A larger number means a larger training time, so more images are shown to the algorithm to learn what a tent is. The problem with Neural Networks is that they can become "overconfident" in their choices. This is caused by the model being trained for too long; training too long will 'saturate' the parameters in the model (neuron saturation). As a result, the model will classify all types of objects as the object we are looking for. However, by training too short, the model will not find tents. Tuning these kinds of parameters is important to get good results with these kinds of models.
Faster R-CNN model
Using a Faster R-CNN model, we were able to identify most tents. So we were able to extract the size of a refugee camp from satellite images. However, because the frequency of satellite imagery on Google Earth is not the same for every location, it also turned out that this source of satellite imagery alone is not enough to implement a model that is accurate enough to make a good estimate of the population in each camp. A collaboration with a satellite image provider could offer a solution here.
Stationsplein 45 A4.004
3013 AK Rotterdam
+31 10 798 62 95