In this thesis, we bridge the gap between artificial intelligence for sustainable science and the inclusivity of computer vision systems. We show how previous approaches to demonstrating the lack of inclusivity of current vision systems have overlooked important aspects of the problem, such as the formalisation of geographical bias and the metrics that reflect its impact. This has led us to propose a protocol for formalising bias, based on the identification of a source, a type and an impact in order to characterise it. This protocol has been implemented for geographical bias, initially on synthetic data, and then experimented with on real data for characterising western bias in vision systems. We find that the results obtained are different from those expected, going against observations in previous academic work. We carry out a visual analysis of these results at different levels of granularity in an attempt to understand them and to propose possible themes for future research. In the end, we highlight the presence of concomitant biases, elements that make up the geographical bias but have different impacts that the main entity. These concomitant biases prevent the characterisation of the geographical bias by influencing the predictions of the models. We therefore show how the problem of characterising geographical bias is more complex than it might at first appear, what the current pitfalls are and what avenues are being pursued to remedy the problems encountered.
Overall, we offer the scientific community tools to better understand the problems of deploying models in developing countries, in order to better understand the challenges of these deployments for applications in sustainable science.