春玉米根系图像语义分割最佳分辨率和概率阈值研究

Optimal Resolution and Probability Threshold for the Semantic Segmentation of Spring Maize Root Image

为了探究根系图像不同概率阈值和分辨率对根系形态参数的影响,以微根管法观测的春玉米(Zea mays L.)抽雄期40 cm深度根系为例,采用降分辨率方法模拟扫描仪获取不同分辨率图像,基于深度学习U-Net语义分割方法建立的iRoot⁃V02根系自动识别软件对不同分辨率根系图像的根长、表面积、体积、直径、根尖数等形态参数进行识别,采用二阶导数方法获取根系与背景的分割阈值拐点,并探讨不同概率阈值对春玉米根系形态参数的影响。结果表明,春玉米抽雄期40 cm深度总根系形态参数(根总长、根总表面积、根总体积、根平均直径、根尖数)和分级根系形态参数(根长、根表面积、根体积)随概率阈值增大呈逐渐减小的趋势,根系形态参数在概率阈值5~20之间的变化幅度较大,在概率阈值25之后趋于稳定;通过曲线拐点方法分析确定,根总长曲线拐点出现在概率阈值25或30处。随着根系图像分辨率的减小,根总长和根尖数呈减小的趋势,而根总表面积和根总体积呈先增大后减小的趋势,根平均直径呈增大的趋势。通过对春玉米抽雄期40 cm深度不同概率阈值根系总的和分级形态参数分析,确定春玉米根系图像识别适宜的概率阈值为25或30;在图像数量较少时采用图像原始分辨率300 dpi分析为宜,图像数量较多时采用215 dpi分辨率为宜。本研究可为春玉米和其他作物根系形态参数的定量研究提供参考。

To explore effects of different probability thresholds and image resolutions on the acquisition of root morphological parameters,root images at 40 cm depth during the tasseling period of spring maize(Zea mays L.)were observed by the minirhizotron method,and were processed by the resolution reduction method to simulate the images with different resolutions obtained by scanners.The root analysis software iRoot-V02 was established by the semantic segmentation method based on the U-Net,a deep learning method,to identify the root length,surface area,volume,diameter and number of root tips from the root images at different resolutions.The inflection point of the segmentation threshold between roots and background is obtained by the second derivative method.Results showed that the total root morphological parameters(total root length,total root surface area,total root volume,average root diameter and number of root tips)and hierarchical root morphological parameters(root length,root surface area and root volume)at 40 cm depth during the tasseling period of spring maize decrease gradually with the increase of probability threshold.The root morphological parameters varied greatly when the probability threshold changed from 5 to 20,but the variation tended to be stable when the probability threshold was greater than 25.It was determined by the curve inflexion point method that the inflexion point of the total root length curve occurred at the probability threshold of 25 or 30.With the decrease of root image resolution,the total root length and the number of root tips decreased,while the total root surface area and the total root volume increased first and then decreased,and the average root diameter increased.Based on the analysis of the total and hierarchical root morphological parameters under different probability thresholds,the appropriate probability threshold for the root image recognition of spring maize was 25 or 30.When the image number is small,it is appropriate to use the original resolution of 300 dpi for the root image analysis,but for mass images the optimal resolution should be reduced to 215 dpi.These findings provide support for quantitative analysis on root morphology of spring maize or other crops.