The agricultural sector is currently facing numerous challenges and structural changes accentuated by demographics, climate change, environmental impact, consumption patterns, competitiveness, among others. To address these challenges, digital technologies, and deep learning in particular, are emerging as one of the main levers. However, as deep neural networks began to be deployed in real-world safety-critical use cases, it became clear that they exhibit a significant gap between their performance “in the lab”, and their predictive quality “in the real world.” This same observation was made during the development of an autonomous spraying system, which must detect and spray herbicide in real time, only where and when needed, and thus must be reliable and trustworthy. This use case is a perfect instance of the problem of decision-making under uncertainty. In order to develop more reliable methods to assess the quality of learning models and provide valid guarantees on the quality of predictions and decisions produced by the system, as well as to exploit the multitude of auxiliary data describing environmental and geographical features associated with images, we turn to conformal prediction as a holistic approach to address these issues. In this context, this thesis offers several contributions to the theory and practical methodology of conformal prediction for the deployment of agricultural robots.
