大豆冠层多光谱图像提取方法

为解决大豆冠层在近地端的多光谱图像边缘灰度不均，目标与背景之间灰度差别小，难以准确高效地获取大豆冠层目标区域的难题，将多光谱成像处理技术与经典图像分割方法有机融合，提出基于多光谱图像处理技术的大豆冠层提取方法。以东北大豆为对象，通过Sequoia多光谱相机采集绿光、近红外、红光、红边和可见光五类大豆多光谱图像，采用高斯平滑滤波法对原始大豆多光谱图像进行预处理，分析多光谱图像中大豆冠层和背景的灰度直方图分布特性，在此基础上利用迭代法、Otsu法和局部阈值法提取原大豆多光谱图像中冠层区域，并以图像形态学开运算处理细化和扩张背景，避免图像区域内干扰噪声对大豆冠层识别效果的影响，同时以有效分割率、过分割率、欠分割率、信息熵以及运行时间等为监督指标，对大豆冠层多光谱图像识别模型进行效果评价。大豆冠层识别模型中迭代法可以有效分割近红外和可见光大豆冠层图像，有效分割率分别为97.81%和87.99%，对绿光、红光和红边大豆冠层图像分割效果较差，有效分割率低于70%；Otsu法和局部阈值法可以有效分割除红光波段的其余四种多光谱大豆冠层图像，且有效分割率均在82%以上；三种算法对红光大豆冠层图像的有效分割率均低于20%，未达到较好效果。在原始多光谱图像中应用迭代法、Otsu法和局部阈值法提取大豆冠层图像与标准图像的信息熵平均值波动幅度分别为：0.120 1，0.054 7和0.059 8，其中Otsu法和局部阈值法较小，表明了对于大豆冠层多光谱图像识别中两种算法的有效性。该算法中Otsu法和局部阈值法均可以有效提取绿光、近红外、红边和可见光等多光谱的大豆冠层图像，二者较为完整地保留了大豆冠层信息，其中Otsu法实时性能较局部阈值法更好。该成果为提取农作物冠层多光谱图像提供理论依据和技术借鉴。

Soybean Canopy Extraction Method Based on Multispectral Image Processing

To solve the problem that the edge gray level of the soybean canopy near the ground is uneven, the gray level difference between the target and the background is small, and it is difficult to accurately and efficiently obtain the soybean canopy target area. This paper combined the multispectral imaging processing technology with the classical image segmentation method and proposed a soybean canopy extraction method based on the multispectral image processing technology. Five kinds of soybean multispectral images, including green light, near-infrared, red light, red edge, and visible light, were collected by Sequoia multispectral camera. A Gaussian smoothing filter preprocessed the original soybean multispectral images. The distribution characteristics of the gray histogram of the soybean canopy and background were analyzed. On this basis, the iterative method, Otsu method, and local threshold method were used to extract the canopy region in the original soybean multispectral image, and the image morphological open operation was used to refine and expand the background to avoid the influence of the interference noise in the image region on the recognition effect of soybean canopy. At the same time, the effective segmentation rate, over-segmentation rate, under-segmentation rate, information entropy, and running time were taken as the monitoring indexes, and the effect of the soybean canopy multispectral image recognition model was evaluated. The results showed that the iterative method could effectively segment the near-infrared and visible soybean canopy images, and the effective segmentation rate was 97.81% and 87.99% respectively. The segmentation effect of green, red and red edge soybean canopy images was poor, and the effective segmentation rate was less than 70%. Otsu and local threshold methods could effectively segment the other four kinds of multispectral soybean canopy images except for red light, and the effective segmentation rate was more than 82%. The effective segmentation rate of the three algorithms for red soybean canopy images was less than 20%, which did not achieve good results. In the original multispectral image, iterative method, Otsu method, and local threshold method were used to extract the mean value of information entropy of soybean canopy image and standard image, and the fluctuation amplitude was 0.120 1, 0.054 7, and 0.059 8, respectively. Otsu and local threshold methods were smaller, showing the effectiveness of the two algorithms in soybean canopy multispectral image recognition. The Otsu and local thresholding methods could effectively extract the soybean canopy images of green light, near-infrared, red edge, and visible light. Both of them retained the soybean canopy information completely. Otsu method had better real-time performance than the local thresholding method. The results provided a theoretical basis and technical reference for extracting crop canopy multispectral images.