Image-Based Estimation of Blueberry Yield Incorporating External Validation and Canopy Architecture Under Field Conditions

Technical Summary: Image-Based Estimation of Blueberry Yield Incorporating External Validation and Canopy Architecture

Problem Statement
Quantifying blueberry fruit yield and maturity is critical for evaluating yield potential in breeding trials and optimizing production logistics. However, manual measurement remains a bottleneck due to being slow, labor-intensive, costly, and prone to subjectivity. While object detection networks offer a high-throughput solution, existing studies often lack rigorous external validation against hand-harvested ground truth, particularly when explicitly accounting for canopy occlusion. Highbush blueberry plants present specific challenges compared to wild varieties due to their complex morphology and significant volumetric occlusion of fruit by foliage. Furthermore, field conditions introduce variability in brightness, photo distance, and perspective, while genotype-specific differences in canopy structure further complicate the visibility of berries to computer vision systems.

Methodology
The study developed a multi-class berry detection pipeline and validated it against hand-harvested counts across 32 diverse southern highbush blueberry genotypes.

• Pipeline Architecture: The system utilizes Slicing Aided Hyper Inference (SAHI) to subdivide high-resolution images into overlapping 640x640px tiles, dynamically adjusting slice size based on image resolution to improve detection of small targets. The detection backbone employs YOLOv8 (specifically comparing nano and extra-large variants) for both single-class and multi-class (immature vs. mature) detection. Post-processing includes non-maximum suppression (NMS) and heuristic filtering based on square aspect ratios to remove erroneous detections.
• Dataset: The "blue-2" dataset comprises 459 annotated images containing 110,220 berry annotations across two maturity classes (immature and mature). The dataset was constructed using images from DSLR cameras and mobile devices, captured via manual and robotic platforms.
• External Validation: Ground-truth validation was conducted on 32 genotypes at Straughn Farms, Florida. Images were captured from a fixed distance (~170 cm) using a Canon EOS Rebel T6 mounted on a rover. All berries were subsequently hand-harvested and counted.
• Canopy Architecture Integration: To address occlusion, the study extracted canopy architecture metrics (e.g., canopy width, height, area, solidity, fractal dimension, texture homogeneity, and color indices) using an OpenCV-based pipeline.
• Predictive Modeling: The relationship between image-based detections, canopy features, and ground-truth counts was modeled using Multiple Linear Regression (LM), Partial Least Squares Regression (PLSR), Ridge Regression, Random Forest, and Support Vector Machines (SVM). Models were evaluated using an 80:20 cross-validation framework repeated 100 times.

Key Results

• Detection Performance: YOLOv8x achieved the highest detection performance with an mAP50 of 0.82 and mAP50–95 of 0.66. The multi-class approach did not significantly degrade detection performance compared to single-class detection. The model demonstrated strong correlation (R² ≥ 0.98) between annotations and inference results.
• Maturity Classification: The model achieved high F1 scores for immature berries (0.91) but lower scores for mature berries (0.74), partly due to color variations during the transition from green to blue and misclassification of intermediate colors.
• Underestimation and Occlusion: Image-based detections systematically underestimated hand-harvested fruit counts, with R² values ranging from 0.40 to 0.61. Hand-harvested counts were approximately four times higher than image-based predictions. This underestimation was attributed to canopy occlusion, which varied widely among genotypes (42% to 90%).
• Impact of Canopy Architecture: Incorporating image-derived canopy architecture features significantly improved prediction accuracy. Partial Least Squares Regression (PLSR) emerged as the best-performing model, increasing R² values for hand-harvested fruit counts from a baseline of 0.52–0.61 to 0.52–0.74 (specifically reaching R² = 0.743 for immature berry counts and 0.74 for maturity ratio).
• Correlations: Canopy architecture features showed moderate associations with occlusion rates (e.g., solidity r=0.49, brightness variability r=0.61), confirming that structural attributes explain variations in fruit visibility.

Key Contributions

1. Validated Pipeline: The study provides a rigorously validated multi-class detection pipeline for highbush blueberries that explicitly accounts for the challenges of small object detection in complex field environments.
2. Quantification of Occlusion: It quantifies the extent of genotype-specific occlusion (42–90%), demonstrating that canopy structure is a primary driver of detection error and that image-based counts alone are insufficient for absolute yield estimation without correction.
3. Integration of Structural Features: The research demonstrates that integrating image-derived canopy architecture (structural, textural, and color features) with fruit detection data significantly improves yield and maturity estimation, offering a method to partially compensate for hidden fruit.
4. Breeding Implications: The study identifies fruit visibility and reduced occlusion as potential selection targets for breeding programs aiming to improve compatibility with high-throughput phenotyping and robotic harvesting.

Significance and Claims
The authors claim that their results demonstrate that accounting for canopy occlusion through the integration of architectural features improves the accuracy of image-based blueberry yield estimation. The study supports the use of computer vision for high-throughput phenotyping but emphasizes that current models tend to underestimate absolute counts due to foliage. By incorporating canopy metrics, the system can better capture biologically meaningful variation among genotypes. The authors conclude that while image-derived phenotypes effectively capture relative differences, future breeding strategies should consider fruit exposure and canopy openness to enhance the accuracy of both phenotyping and robotic harvesting systems. The paper does not claim to have solved the occlusion problem entirely but rather provides a methodological framework to mitigate its impact on yield estimation.