基于K­means聚类和分区寻优的秸秆覆盖率计算方法

Corn Straw Coverage Calculation Algorithm Based on K­means Clustering and Zoning Optimization Method

摘要:

针对农田秸秆形态多样、细碎秸秆难以准确识别的问题，基于机器视觉技术，提出了一种基于K-means聚类和分区寻优结合的秸秆覆盖率计算方法。该方法首先利用K-means聚类算法对玉米秸秆图像进行分割，使秸秆从背景图像中分离；然后将秸秆图像分隔为16区，利用统计学方法分别计算各区秸秆中位数和众数灰度平均值，16区平均后分别获得秸秆中心灰度和土壤背景中心灰度，将其作为新的分类中心，重新采用K-means聚类方法对玉米秸秆图像进行分割，当秸秆中心灰度不再发生变化时停止迭代，计算秸秆像素点数量；最后计算获得玉米秸秆覆盖率。2021年4月，该方法在吉林省长春市玉米地100个采样点进行了实际验证，与人工拉绳法和人工图像标记法的相关系数分别为0.7161和0.9068，误判率7%，平均误差比Otsu阈值化方法和经典K-means聚类方法分别降低了45.6%和29.2%。试验结果表明，所提方法能够实现对不同天气、不同种植模式、不同地块条件下的秸秆覆盖率准确计算，该研究结果可为秸秆覆盖率在线计算提供一种新方法。

关键词:玉米秸秆覆盖率;机器视觉;K­means聚类;分区寻优;保护性耕作

Abstract:

Straw coverage rate is one of the most important indicators for conservation tillage evaluation. It is also important to realize online detection of straw coverage rate for black land conservation evaluation. Aiming at the problems of various forms of cropland straws and the difficulty in accurately identifying the broken straws, an online detection algorithm of straw coverage rate was proposed based on the combination of K-means clustering and zoning optimization method with machine vision technology. Firstly, K-means clustering algorithm was used for maize straw image segmentation from the background image. And then the straw image was segmented into 16 areas, using statistical methods to calculate respectively the median of straw and the average number of gray levels, respectively. After 16 area average straw center gray value and soil background gray value were calculated, both were taken as a new classification center. Subsequently the K-means clustering method was used to segment the image of corn straw again. When the gray value of the center of straw did not change, the iteration was stopped. With the calculated number of pixel points of straw, the coverage rate of corn straw was obtained, finally. In April 2021, the proposed algorithm was verified at 100 sampling points in corn fields in Changchun City, Jilin Province. The correlation coefficients between the proposed algorithm and the artificial rope pulling method and the artificial image labeling method were 0.7161 and 0.9068, respectively. The corn straw coverage detection misjudgment rate was 7%. The average error of Otsu thresholding method and classical K-means clustering method was respectively reduced by 45.6% and 29.2%. The experimental results showed that the proposed method could detect the straw coverage rate under different weather conditions and planting patterns, accurately. It can provide a method for online detection of straw coverage.

Key words:corn straw coverage; machine vision; K­means clustering; zoning optimization; conservation tillage