Description

Drones now provide imagery offering unprecedented detail and machine learning can translate such images into information at remarkable accuracy. Machine Learning is expected to support the monitoring and representation of several Sustainable Development Goal indicators, especially in low-income countries. Yet society is wary of adopting these technologies, since complex and uninterpretable results can affect the accountability of decision- and policy-makers wishing to benefit from machine learning.
Supervised machine learning methods face two interrelated challenges. First, target classes are highly variable and are usually embedded in a heterogeneous natural or anthropogenic landscape. Second, available training samples are not enough to adequately represent this high level of variability characterizing the target classes. These challenges are the main reason why supervised methods work efficiently on a local scale, but provide less reliable classification results when reused across different regions. Therefore, innovative solutions are required to guide generalization of supervised methods in the spatial domain, make machine learning methods more transparent, help explain the results to decision- and policy-makers, and guide the machine learning engineer to label samples which will have the greatest influence in improving the results.

Objectives and Methodology

This study will make use of labelled UAV / satellite imagery across different urban environments. A supervised machine learning method (e.g. Fully Convolutional Networks) will be trained on one study area, and the generalization of that model will be assessed on different study areas. Improvements will be made to improve the transparency of the results to give end-users a better idea of the accuracy of the predictions. Additionally, active learning methods will be employed to indicate which samples should next be labelled in order to improve the results when generalized to a different domain.

Further reading