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

Abstract: 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.