玉米作物多光谱图像精准分割与叶绿素诊断方法研究

为了快速获取大田玉米作物长势信息，基于多光谱图像开展了大田玉米叶绿素指标的非破坏性诊断研究。应用自主开发的2-CCD多光谱图像感知系统，在田间采集玉米冠层可见光[Blue(B)，Green(G)，Red(R);400～700 nm]和近红外(Near-infrared: NIR，760～1 000 nm)图像，并使用SPAD同步测量样本叶绿素指标。采集后图像经自适应平滑滤波处理后，进行图像玉米植株提取。为了选择最优算法实现玉米植株与杂草、土壤背景的分割，首先比较了最大类间方差(OTSU)分割算法和局部阈值处理分割算法，选取了基于局部统计的可变阈值处理方法对玉米NIR图像进行初步分割，进而采用区域标记算法进行精细分割，分割准确率达95.59%。将分割结果应用于玉米植株可见光图像R，G，B各通道，从而实现了玉米植株多光谱图像中可见光图像的整体分割。基于分割后R，G，B和NIR四个通道的玉米冠层图像，提取了各通道图像灰度均值(A_NIR，A_Red，A_Green和A_Blue)并计算了归一化植被指数(NDVI)、比值植被指数(RVI)和绿色归一化植被指数(NDGI)作为光谱特征参数，建立了玉米冠层叶绿素指标诊断的偏最小二乘法回归模型。结果表明，建模R2达0.596 0，预测R2达0.568 5，该方法通过玉米多光谱图像特征参数评估叶片叶绿素含量，可为大田玉米长势监测提供支持。

Research on Maize Multispectral Image Accurate Segmentation and Chlorophyll Index Estimation

In order to rapidly acquire maize growing information in the field, a non-destructive method of maize chlorophyll content index measurement was conducted based on multi-spectral imaging technique and imaging processing technology. The experiment was conducted at Yangling in Shaanxi province of China and the crop was Zheng-dan 958 planted in about 1 000 m×600 m experiment field. Firstly, a 2-CCD multi-spectral image monitoring system was available to acquire the canopy images. The system was based on a dichroic prism, allowing precise separation of the visible (Blue (B), Green (G), Red (R): 400～700 nm) and near-infrared (NIR, 760～1 000 nm) band. The multispectral images were output as RGB and NIR images via the system vertically fixed to the ground with vertical distance of 2 m and angular field of 50°. SPAD index of each sample was measured synchronously to show the chlorophyll content index. Secondly, after the image smoothing using adaptive smooth filtering algorithm, the NIR maize image was selected to segment the maize leaves from background, because there was a big difference showed in gray histogram between plant and soil background. The NIR image segmentation algorithm was conducted following steps of preliminary and accuracy segmentation: (1) The results of OTSU image segmentation method and the variable threshold algorithm were discussed. It was revealed that the latter was better one in corn plant and weed segmentation. As a result, the variable threshold algorithm based on local statistics was selected for the preliminary image segmentation. The expansion and corrosion were used to optimize the segmented image. (2) The region labeling algorithm was used to segment corn plants from soil and weed background with an accuracy of 95.59%. And then, the multi-spectral image of maize canopy was accurately segmented in R, G and B band separately. Thirdly, the image parameters were abstracted based on the segmented visible and NIR images. The average gray value of each channel was calculated including red (A_Red), green (A_Green), blue (A_Blue), and near-infrared (A_NIR). Meanwhile, the vegetation indices (NDVI (normalized difference vegetation index), RVI (ratio vegetation index), and NDGI(normalized difference green index)) which are widely used in remote sensing were applied. The chlorophyll index detecting model based on partial least squares regression method (PLSR) was built with detecting R2=0.596 0 and predicting R2=0.568 5. It was feasible to diagnose chlorophyll index of maize based on multi-spectral images.