基于玉米冠层原位监测的全生育期叶色建模及其应用

针对田间玉米冠层叶色变化难以定量描述问题,该文利用田问原位冠层监测系统,在摄像机自动曝光模式下连续采集多个玉米品种的冠层图像,揭示了复杂天气条件对图像和玉米冠层颜色的影响.利用概率密度统计分析方法分别计算玉米6个关键生育期的冠层亮度-色度分布,并针对冠层色度具有明确变化趋势且分离度较高的冠层亮度区间,建立了全生育期玉米冠层叶色模型.进而,基于该模型建立了适合不同玉米生育期的冠层图像自动分割方法,将玉米全生育期的冠层图像分割精度提升到82.6％,并揭示了不同品种玉米在叶片发育过程中冠层叶色与叶龄的相关性,利用登海605和农大108的冠层叶色预测出的生育期叶龄均方根误差RMSE (root mean squared error,RMSE)分别为1.14和1.41叶.试验结果表明,该文建立的玉米冠层叶色模型能够较好描述玉米关键生育期的冠层叶色变化规律,对玉米冠层图像分割、生育期估计、玉米品种表型鉴定具有重要意义.

Modeling of maize canopy color in whole growth period based on in-situ monitoring system and its application

Abstract: Maize canopy leaf color is an intuitive reflection of maize growth, development, and physiological and biochemical status, and also an important trait for maize phenotypic detection in the field investigation. Both visual observation and quantitative analysis showed that different solar radiation had significant effects on maize canopy hue (CH), and the CH value of maize canopy had significant changes at different growth stages. Nowadays, high-throughput phenotyping platforms have gradually been applied from controllable indoor environment to uncontrollable field environment, however, the complex field condition brings a lot of challenges to the current phenotyping techniques. In the field-based maize growth monitoring application, how to quantitatively analyze the color variation tendency of maize canopy in field environment is still an urgent problem to be solved. In this study, we developed sets of in-situ monitoring systems in the field to continuously capture canopy image sequences for 2 maize cultivars (DH 605 and ND 108) in the whole growth stage, and respectively collected 6 data sets of maize canopy image in consideration of 2 types of different weather conditions (sunny and cloudy days) and 6 key growth stages (4 leaves, 9 leaves, 16 leaves, silk, blister and milk stages). These image data sets of maize canopy not only reflected the effect of different weather conditions on canopy color, but also reflected the natural changes of canopy color at different growth stages, so they could be used for the color quantification and evaluation of maize canopy. With these data sets, statistical analysis based on the HSV (hue, saturation, value) color space in the pixel level was utilized to reveal the relationship among solar radiation, image color and canopy color. The results of quantitative analysis showed: Solar radiation had little effect on image value (IV) and CH, but had great effect on image hue (IH) and canopy value (CV), and the distribution of the canopy pixels at the same CV value was approximately consistent with the normal distribution. And then, the canopy CV-CH distributions of 6 key growth stages of maize were calculated respectively by probability density statistical techniques. These distributions manifested clear variation tendency and distinction degree in CV domain from 80 to 200, which meant that the CH statistical values in this CV domain could be used to quantify and evaluate color differences among various growth stages of maize. Therefore, a continuous maize canopy color model (MCCM) was established based on the statistical results of 6 key growth stages, which described the successive color change of maize canopy in the whole growth stage. During the stage from leaf emergence to development (4 leaves, 9 leaves and 16 leaves stages), the CH values of maize showed a significant decreasing trend, and then the CH values increased gradually in the silk, blister and milk stages. Based on this model and CV-CH distribution, maize canopy segmentation method was further designed for different growth stages and field conditions. By the comparison with other segmentation methods based on color indices, such as color index of vegetation extraction (CIVE), excess green (ExG), excess green-excess red (ExGR), vegetation (VEG) and hue (H), the presented method could effectively improve the canopy segmentation accuracy, and obtain a segmentation accuracy of over 82.6% for maize canopy images in the whole growth stage. At the same time, this model revealed a significant correlation between the CH value and emerged leaf number (ELN) of 2 maize cultivars (i.e. Denghai605 and Nongda108), and the RMSE (root mean square error) values were 1.14 and 1.41 leaves respectively. The experimental results demonstrate that the maize canopy color model can quantitatively describe canopy color variation in different maize stages, and has important application value for the automatic image segmentation of maize canopy, the prediction of growth stages, and the phenotype identification of maize cultivars.