Abstract:

Minirhizotron technique can directly monitor the dynamic growth and development of plant roots and be used to obtain clear root images. However， because of the complex soil environment， uneven particle size and large number of fine roots， it easily causes discontinuity of the divided roots and mistakes the soil background as the root. To solve the above problems， the CP-DeepLabv3+ algorithm was proposed to segment image. The coordinate attention mechanism （CA） was introduced to effectively segment the target location information and made the edge of the target more continuous. Strip pooling （SP） branch was added to ASPP feature extraction module to avoid unnecessary connections between distant locations and improve the accuracy of image segmentation. CP-DeepLabv3+ algorithm was applied to test maize root dataset. The results showed that the mean intersection-over-union （MIoU） value was 82.95%， the mean pixel accuracy （MPA） value was 92.47%， which was 3.69% and 4.44% higher than the original DeepLabv3+ model， respectively. This algorithm could effectively segment maize root and has practical significance for feature extraction.

Key words: maize root system, microroot canal method, in situ monitoring, CP-DeepLabv3+

摘要：

利用微根管技术可以直接监测植物根系动态生长，并获取清晰根系图像，但土壤环境复杂、颗粒不均匀、细根数量多，图像分割时容易造成根系不连续，将土壤背景误认为根系。针对以上问题，提出了CP-DeepLabv3+算法进行图像分割。该算法引入坐标注意力机制（coordinate attention， CA），更精确地获得分割目标信息，使得分割目标边缘更加连续；在ASPP特征提取模块加入条纹池化（strip pooling，SP）分支，避免在相距较远的位置之间建立不必要的连接，提高图像分割精度。利用CP-DeepLabv3+算法对玉米根系数据集进行测试，结果显示，平均交并比（mean intersection over union，MIoU）值为82.95%，平均像素精确度（mean pixel accuracy，MPA）值为92.47%，相比于原始DeepLabv3+模型分别提高了3.69%、4.44%，表明该算法可有效分割玉米根系，对图像特征提取具有实际意义。

关键词: 玉米根系, 微根管法, 原位监测, CP-DeepLabv3+