Zhang Weirong, Chen Xuegeng, Qi Jiangtao, Yang Sisi
Key Laboratory of Bionic Engineering, Ministry of Education, Jilin University, Changchun, China.
College of Biological and Agricultural Engineering, Jilin University, Changchun, China.
Front Plant Sci. 2022 Dec 1;13:1041791. doi: 10.3389/fpls.2022.1041791. eCollection 2022.
The widespread use of unmanned aerial vehicles (UAV) is significant for the effective management of orchards in the context of precision agriculture. To reduce the traditional mode of continuous spraying, variable target spraying machines require detailed information about tree canopy. Although deep learning methods have been widely used in the fields of identifying individual trees, there are still phenomena of branches extending and shadows preventing segmenting edges of tree canopy precisely. Hence, a methodology (MPAPR R-CNN) for the high-precision segment method of apple trees in high-density cultivation orchards by low-altitude visible light images captured is proposed. Mask R-CNN with a path augmentation feature pyramid network (PAFPN) and PointRend algorithm was used as the base segmentation algorithm to output the precise boundaries of the apple tree canopy, which addresses the over- and under-sampling issues encountered in the pixel labeling tasks. The proposed method was tested on another miniature map of the orchard. The average precision (AP) was selected to evaluate the metric of the proposed model. The results showed that with the help of training with the PAFPN and PointRend backbone head that AP_seg and AP_box score improved by 8.96% and 8.37%, respectively. It can be concluded that our algorithm could better capture features of the canopy edges, it could improve the accuracy of the edges of canopy segmentation results.
无人机（UAV）的广泛应用对于精准农业背景下果园的有效管理具有重要意义。为减少传统的连续喷洒模式，可变目标喷雾机需要有关树冠的详细信息。尽管深度学习方法已在识别单株树木领域广泛应用，但仍存在树枝延伸和阴影现象，导致难以精确分割树冠边缘。因此，提出了一种通过捕获的低空可见光图像对高密度栽培果园中的苹果树进行高精度分割的方法（MPAPR R-CNN）。采用具有路径增强特征金字塔网络（PAFPN）和PointRend算法的Mask R-CNN作为基础分割算法，以输出苹果树树冠的精确边界，解决了像素标注任务中遇到的过采样和欠采样问题。所提出的方法在果园的另一张小地图上进行了测试。选择平均精度（AP）来评估所提出模型的指标。结果表明，借助PAFPN和PointRend主干头进行训练，AP_seg和AP_box得分分别提高了8.96%和8.37%。可以得出结论，我们的算法能够更好地捕捉树冠边缘特征，提高树冠分割结果边缘的准确性。
