Synthetic Data-Driven AI Using Mixture of Rendered and Real Imaging Data for Strawberry Yield Estimation

The use of synthetic data in agriculture can benefit researchers and agricultural business innovators by reducing the amount of time, effort, and cost required to collect field data. Current approaches for collecting synthetic image data in agriculture applications rely on sample images found on the web or attempting to recreate an agricultural scene that does not accurately reflect the targeted environment. In this study, images were collected from a strawberry farm using cameras mounted on a ground vehicle which traversed over plasticulture strawberry bed rows. The strawberry farm was then recreated in simulation with 3D models of strawberry plants to be used for capturing synthetic 2D image data. These datasets were used to train three Faster R-CNN neural networks for fruit detection and applied to the strawberry farm field images for inference. The synthetic images, despite having less structural variability than the real images, were still effective in identifying fruit in test images from the field. Strawberry fruit detection accuracy was 68% (F1-score 0.80) using only synthetic data and 90% (F1-score 0.93) using a mixture of synthetic and real images, showing that synthetic data can be used to reduce the number of real images needed for training an object detection neural network model. The study demonstrates the potential for synthetic data in agricultural technology development, particularly for fruit detection. However, future work is necessary to further investigate features needed to refine synthetic crop models for improved detection accuracy in field conditions.