Deep Learning and Topology Analysis for Pruning Target Decision and Pruning Point Localization in Dormant Walnut Trees

Accurate pruning target decision and pruning point localization are essential for intelligent walnut pruning but remain challenging because slender and crossing branches are difficult to segment during the dormant season. This study proposes an improved semantic segmentation method combined with a three-stage progressive rule inference strategy for pruning target decision and pruning point localization. SAA-TransUNet was developed by incorporating an enhanced atrous spatial pyramid pooling module (EABP-ASPP) and a context-aware gated feature fusion module (CAGF) to improve trunk and branch segmentation. The predicted masks were skeletonized to construct branch topology based on endpoints, bifurcation points, and branch segments. Geometric and topological features were then used to sequentially determine pruning targets for upright, inward-growing, and competing branches and localize their pruning points. SAA-TransUNet achieved an mIoU of 84.93% and an IoUbranch of 80.41%, with the latter increasing by 13.05 percentage points over the original TransUNet. On the independent test set, the pruning target decision achieved an overall precision of 86.34%, with a mean pruning point localization error of 3.52 pixels and an average processing time of 2.154 s per image. The proposed method integrates branch segmentation, topology analysis, pruning target decision, and pruning point localization, providing a basis for subsequent robotic pruning operations in dormant walnut trees.