In order to perform online and nondestructive measurements of the parameters of flora, the use of machine vision technology was investigated. This technology was used to capture the image of a flora canopy, and then three segmentation algorithms: Excess Green (ExG) minus Excess Red (ExR), ExG, and normalized difference indices (NDI) were used to extract the canopy area of the flora. The ExG and NDI used an Otsu threshold value to obtain a binary image, and the ExG-ExR used a fixed threshold value to obtain a binary image. Flora canopy characteristic parameters (covering ratio, canopy length, and canopy width) were extracted based on the projection profile of the canopy leaves extracted by the flora canopy segmentation methods. These were combined with the parameters of the flora obtained by artificial measurement: stem height, stem diameter, leaf number, fruit number, and LAI (fitting value), to form five types of inversion models for the five growth parameters of the flora. The inversion models were based on the covering ratio, canopy width, and canopy length, and a regression equation established by three parameters of the flora and an average inversion model were established. The results showed that the contact ratio and recognition rate of extraction of the flora canopy region, using the segmentation method ExG-ExR, were more than 99.5% and 98.2%, respectively. Furthermore, identification of the flora canopy was accurate, and there were very few mistakenly identified areas. No matter when the image was captured, the recognition performance of the flora canopy image was stable, and the performance was superior to the methods ExG+Otsu and NDI+Otsu. The contact ratio of the ExG+Otsu segmentation method ranged from 72.7% to 93.5% and recognition rate was 71.1%-90.2%, and showed a small amount of leakage and error used to partition the flora canopy figures. The contact ratio of the NDI+Otsu segmentation method ranged from 99.9% to 100%, however, the scope of the recognition rate was 13.1%-89.2%, and showed a high incidence of false recognition and unstable performance. Inversion models were validated using 120 new images. The inversion results showed that the regression coefficient between the inversion value and the measured value was greater than 0.958 when using the inversion model of the flora canopy covering ratio. The performance of the flora canopy covering ratio was superior to the inversion models of canopy width and canopy length. The inversion model using the regression equation and the average model were the only two parameters that were better than the inversion model of the covering ratio. Between the inversion values of stem height, stem diameter, leaf number, fruit number, LAI, and the measured values of each, the regression coefficient were 0.979, 0.976, 0.979, 0.965, and 0.973, respectively, and the SEwere 10.55 cm, 1.37, 0.213 mm, 0.672, and 0.055, respectively. The inversion method, based on machine vision technology, can achieve online and nondestructive measurements of the parameters of flora, which can provide significant advances in controlling the greenhouse environment and precise fertigation.%为实现植物群生长参数在线无损检测，采用机器视觉技术捕获植物群冠层图像，通过RGB空间超绿色-超红色指标（excess green minus excess red，ExG-ExR）、超绿色指标（excess green，ExG）和归一化差异指标（normalized difference indices，NDI）3种指标分割植物群冠层图像，提取植物群图像特征参数：覆盖率、冠层幅长和冠层幅宽，并结合人工测量植物群体参数：茎秆高度、茎直径、叶面数量、坐果数量和叶面指数（leaf area index，LAI）（拟合值），建立植物群5个生长参数的5种反演模型分别为覆盖率反演模型、冠层幅宽反演模型、冠层幅长反演模型、回归方程反演模型和均值反演模型。结果表明：采用 ExG−ExR 分割的植物群冠层区域与人工提取区域重合度大于99.5%，识别率大于98.2%，分割性能优于 ExG+Otsu 和 NDI+Otsu 分割方法。采用120幅反演模型验证图验证反演模型性能，结果表明植物群冠层覆盖率反演模型反演5个植物群生长参数时，其反演值与测量值间相关性决定系数大于0.958，性能优于冠层幅宽和幅长反演模型，而回归方程和均值反演模型在反演植物群5个生长参数时，都仅有2个参数反演性能优于覆盖率反演模型。茎秆高度、叶面数量、茎直径、坐果数量和LAI的反演模型反演值与测量值间线性相关决定系数最高分别为0.979、0.976、0.979、0.965和0.973，标准误差（standard error，SE）分别为10.55 cm、1.37、0.213 mm、0.672和0.055，其对应的反演模型分别为均值反演模型、覆盖率反演模型、覆盖率反演模型、覆盖率反演模型和均值反演模型。通过机器视觉技术及反演模型能够在线无损准确反演植物群生长参数，为温室环境调控及精准肥水一体灌溉控制系统提供具有代表性意义的决策依据。
展开▼
