Early Season Crop Load Estimation and Yield Prediction in Apple Orchards
Open Access
- Author:
- Jarvinen, Thomas Dunn
- Graduate Program:
- Agricultural and Biological Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 01, 2019
- Committee Members:
- Daeun Choi, Thesis Advisor/Co-Advisor
Paul Heinz Heinemann, Committee Member
Tara Auxt Baugher, Committee Member
James Rawlinson Schupp, Committee Member - Keywords:
- Precision Agriculture
Computer Vision
Yield Prediction
Yield Mapping
Deep Learning
Multiple Object Tracking - Abstract:
- In apple production, early season yield predictions have applications in labor demand projections, packing-house inventory management, and sales planning. To make these predictions, fruit counts are a useful metric, but it is expensive to measure these by hand. To automate this process, a computer vision system using a Faster R-CNN object detector was developed which detected immature fruit with a precision of 0.85 and a recall of 0.92, and detected mature fruit with a precision of 0.92 and a recall of 0.82. However, for the purpose of counting fruit even a highly accurate detector will fail to find many apples due to occlusion. It has been proposed that analyzing video streams rather than single images could minimize accuracy loss due to occlusion. To this end, the CNN based object detector was integrated with a video multiple-object tracking algorithm to produce fruit counts for early-season yield prediction. This system was demonstrated to be an accurate fruit counting system which is robust to variability in fruit maturity, and tree structure. Video based fruit counting was found to be a strong predictor of yield, predicting with R 2 = 0.81 when compared against harvest weight. The resulting yield prediction system was tested on trees with varying canopy depths to find a relationship between canopy density and prediction accuracy. No significant difference was found in prediction accuracy over the different pruning severites. The resulting system has potential applications as an early season yield predictor in a commercial setting.