Abstract:
Data relating to socio-economic and environmental conditions is scarce in the developing world, which leads to a lack of reliable information on which to base policy decisions on. This is one of the biggest challenges in fighting issues such as poverty and deforestation. Government, Non-Government Organizations, and other international organizations often fill in the gaps with door-to-door surveys, but these can be expensive and time-consuming to conduct. Research suggests that information about poverty and forestation can be extracted from high-resolution satellite imagery using machine learning and image processing techniques with considerably accurate results.
In this project, we have used machine learning models to extract information from satellite imagery that has helped map development, and forestation has been mapped using image processing techniques.
For the poverty mapping module, we used a pre-trained Convolutional Neural Network Model, modified to act as feature extractor. The output feature vectors from this model, in addition with ground truth obtained from survey data, were used to further train a regression model. The results from this trained regression model represent poverty scores. This method has proved quite successful with an error rate of around 10%.
Our forestation mapping follows a simpler image processing technique. With the assumption that green parts in a satellite image represent green areas on ground, we aimed to segment these areas, and take them as a measure of forestation. However, it cannot be distinguished whether the green patch belongs to a forest or a field. Similarly, individual trees which look like rocks from an aerial view might also be missed if viewed from too high an altitude. However for general trend mapping, our technique has shown good results.
