An automated vine pruning robot is being developed to reduce the cost of labour in vineyards. This automated system requires an accurate model of the vine’s structure, including the locations of support wires, in order for the robot to make good decisions about where and how to prune the plant. In this project we have developed a system for accurately and robustly detecting pixels belonging to wires in Bayer Images taken by the robot of the vine’s canopy. Our system uses support vector machines for classifying wire and non-wire pixels, and a set of masks for optimally distributing training examples over an image. We have found an optimal subset of features for describing these examples and are able to achieve upwards of 90% precision with more than 20% recall. The system generates data ideal for wire fitting and use by the automated vine pruning robot. The techniques discussed could be generalised and used in other scenarios where selecting ideal example data from a large pool of potential examples, and finding optimal features to represent these examples is required.