Bounded transparency for automated inspection in agriculture

In agriculture, a major challenge is to automate knowledge-intensive tasks. Task-performing software is often opaque, which has a negative impact on a system’s adaptability and on the end user’s understanding and trust of the system’s operation. A more transparent, declarative way of specifying the expert knowledge required in such software is needed.We argue that a white-box approach is in principle preferred over systems in which the applied expertise is hidden in the system code. Internal transparency makes it easier to adapt the system to new conditions and to diagnose faulty behaviour. At the same time, explicitness comes at a price and is always bounded by practical considerations. Therefore we introduce the notion of bounded transparency, implying a balanced decision between transparency and opaqueness. The method proposed in this paper provides a set of pragmatic objectives and decision criteria to decide on each level of a task’s decomposition whether more transparency is sensible or whether delegation to a black-box component is acceptable.We apply the proposed method in a real-world case study involving a computer vision application for seedling inspection in horticulture and show how bounded transparency is obtained. We conclude that the proposed method offers structure to the application designer in making substantiated implementation decisions.