基于双侧图像识别的单株苹果树产量估测模型

Yield estimation model of single tree of Fuji apples based on bilateral image identification

利用普通数码相机获取成熟期苹果树图像进行产量估测,具有成本低、操作简单等特点,其关键是估测模型的建立。该文分别按东南和西北2个方向获取富士苹果成熟期的40株果树的80幅图像,通过果实特征提取,获取东南方向识别出的图斑数量(参数1)、西北方向识别出的图斑数量(参数2)、东南方向识别出的图斑像素面积(参数3)、西北识别出的图斑像素面积(参数4),分别以识别出的4个参数及双方向图斑数量之和(参数5)、双方向图斑像素面积之和(参数6)共6个参数为自变量,以获取的单株产量信息为因变量,以奇数组20株果树为建模数据集建立线性回归模型。结果表明以参数5构建的产量估测模型的决定系数R2最高为0.81,相对均方根差(NRMSE)值最低为0.11,说明以该参数构建的模型其估测效果最好;进一步利用以参数5构建的估测模型对偶数组20株果树进行验证,其NRMSE值为0.16,估测结果较好,但也存在估测产量较大波动的情况。深入讨论引起估测偏差的情况,后期研究应重点提高逆光、弱光照条件下的成熟期苹果的识别率,及解决由于单果因遮挡被分离而被识别为多果的情况和多果因重叠被识别为单果的情况,以提高识别效果,进而提高产量模型估测效果。

Apples yield estimation with a common digital camera to get mature fruits, has the advantages of low lost, simple operation and other characteristics. Key to the estimation is the establishment of an estimation model. In this paper, 80 images from 40 Fuji trees were acquired from the southeast and northwest directions using a Cannon G7 camera. By fruit feature extraction, 4 parameters were identified, which were identification patch number from southeast direction （parameter 1）, identification patch number from northwest direction （parameter 2）, patch pixels area from southeast direction （parameter 3） and patch pixels area from northwest direction （parameter 4）. A total of 6 parameters, including the above-mention 4 parameters along with the sum of patch number from two directions （parameter 5） and the sum of patch pixels area from two directions （parameter 6） acted as independent variables and single tree yield information acted as the dependent variable. With 20 fruit trees used as the modeling data set, the linear regression model was constructed based on the independent variables and dependent variable. The results showed that the yield estimation model with parameter 5 had the best effects with the highest R2 of 0.81 and the lowest NRMSE （Normal Root Mean Squared Error） value of 0.43. Further, additional 20 fruit trees were verified using the yield estimation model with parameter 5. The estimation result was good with a NRMSE value of 0.59, but there were also fluctuations between estimation yield and actual yield. In the verified 20 fruit trees, there were 10 trees whose estimated yield was higher than the actual yield, and the deviation value of No. 2 tree was maximum of 14.02. There were also 10 trees whose estimated yield was lower than the actual yield, and the deviation value of No. 30 was maximum of 17.79. The reason of estimation errors was discussed. Later studies should focus on improving mature apple recognition rates in conditions of backlighting and weak light, and solve the error recognition in conditions of single apple occlusion and multi apples overlapping. The research will help improve recognition effects and then improve model estimation effects.