Fairness and Interpretability for Location Quality prediction from Optical Images

Computer vision datasets are increasingly being used to predict the qualities of places, such as the aesthetic quality of landscapes or the liveability of cities, where increasingly challenging and complex tasks are carried out to understand location qualities. As a result, location quality assessments can be performed at an increasingly grand scale and with greater accuracy. However, comparatively little attention has been given to finding methods to understanding both the datasets and the models used to predict liveability. Datasets have latent biases and are poorly-understood in what the images of the dataset contains. Meanwhile, models offer limited options to understand how predictions are made. To improve our understanding of location qualities, we must address these issues so that we can create systems which allow for dependable knowledge extractions. In this thesis I will contribute to the availability of methods which contribute to the fairness and the interpretability of location quality prediction from optical images, both remotely-sensed images as well as geotagged photographs.