Fine-tuning and testing of a deep learning algorithm for pruning regions detection in spur-pruned grapevines

Vineyard management cost is significantly affected by selective operations such as winter pruning. This study aimed to fine-tune and test a deep neural network (DNN) based algorithm for detecting pruning regions. Focusing on spur-pruned grapevines, in order to fine tune the DNN, around 1000 RGB images were acquired and pruning target regions ground-truthed. The DNN was tested on 5 vines, 232 frames were acquired and processed in real-time for identifying the regions of interest as potential pruning regions (PPRs). PPRs were then classified depending on wood type, orientation and visibility. True positives (TPs), false positives (FPs) and false negatives (FNs) were identified in each frame. Best detection performance was obtained for visible coplanar simple spurs (recall = 98%) while the recall index was generally lower than 60% when pruning regions (PRs) were not clearly visible. FPs were more frequently associated with old cuts located on permanent organs such as trunks and cordons.