"We present a new approach towards the automation of cane pruning of wine grape vines, a challenging task crucial for good yield and quality. We recognised that expert pruners assess the entire vine architecture when pruning, and do not solely rely on the simple pruning rules given by viticulturists. Rather, custom pruning behaviour that is sometimes unique to the vineyard, and may be considered part of their brand, is common. We therefore developed this model that represents the vine architecture as a mathematical graph and learns from real pruning examples, provided by experienced pruners, using a Graph Neural Network (GNN).

"This new model improves upon previous research by learning from realistic vine data and expert pruning examples, moving away from synthetic data and the flawed assumption of a single perfect pruning solution. Instead, the GNN is trained to accurately score the quality of multiple potential pruning solutions."