Abstract
Precision agriculture requires efficient crop yield estimation, yet counting small, dense fruits like cranberries in aerial imagery is challenging. Manual pixel-wise annotation is prohibitively expensive. This paper introduces the Triple-S Network, a deep learning framework that achieves high-accuracy segmentation and counting using only point-click supervision. By incorporating geometric shape priors (circularity and convexity) and a selective watershed algorithm, the method improves counting accuracy by nearly 23% over state-of-the-art baselines on the new Cranberry Aerial Imagery Dataset (CRAID).
Introduction
Estimating crop yield is vital for agricultural resource management and profit prediction. While Unmanned Aerial Vehicles (UAVs) provide a non-invasive, high-coverage alternative to manual sampling, processing the resulting imagery remains difficult. Cranberries present specific challenges: they are small, numerous, and often obscured by foliage. Furthermore, late-season red leaves act as distractors for color-based algorithms. Training robust models typically requires expensive pixel-wise masks. This research addresses whether high-quality segmentation and counting can be achieved using only a single "point-click" at the center of each berry.
Methods: The Triple-S Network
The proposed Triple-S Network (Segmentation, Split, and Shape) utilizes a U-Net backbone with three specialized loss branches to overcome the limitations of weak point supervision.
- Segmentation Branch: Learns the general spatial distribution of berries from point annotations.
- Split Branch: Addresses the issue of overlapping berries. A Selective Watershed algorithm uses both berry centers (positive seeds) and background points (negative seeds) to delineate boundaries between touching objects, preventing them from being merged into single blobs.
- Shape Branch: Since point labels lack geometric information, the authors introduce Shape Priors. By enforcing constraints for convexity and circularity in the loss function, the network is regularized to predict elliptical shapes that match the physical geometry of cranberries.
The Cranberry Aerial Imagery Dataset (CRAID)
To validate the approach, the researchers introduced CRAID, the largest dataset of its kind, featuring 21,436 drone-captured images. The dataset covers the full growth cycle, including the challenging "red leaf" period. While the bulk of the data uses low-cost point annotations, a subset of 702 images includes full pixel-wise masks for rigorous evaluation.
Results and Discussion
The Triple-S Network significantly outperformed existing baselines like standard U-Net and LC-FCN.
Key performance metrics include:
- Counting Accuracy: A 22.91% improvement in Mean Absolute Error (MAE) over previous state-of-the-art methods.
- Segmentation Quality: An increase in Mean Intersection over Union (mIoU) by over 6.74%.
- Robustness: Qualitative analysis confirms the model successfully distinguishes berries from red leaves by leveraging learned shape and contextual cues.
Conclusion
The Triple-S Network demonstrates that point-wise supervision, when combined with geometric priors, is a viable and cost-effective alternative to full pixel-wise annotation for agricultural monitoring. This approach reduces annotation costs by approximately 10x while improving both counting and segmentation accuracy. Future work may extend these shape-constrained models to other round crops such as blueberries or olives.
Summary updated