ÌÇÐÄÊÓƵ

A research team demonstrated that the Segment Anything Model (SAM) accurately identifies individual berries in 2D grape cluster images, achieving a strong correlation with human-identified berries (Pearson's R² = 0.96). This method generated over 150,000 berry masks from approximately 3,500 images.

The results show the potential for integrating SAM into existing vineyard image-processing pipelines to improve cluster architecture and compactness analysis. Future applications include enhanced vineyard management and breeding practices by providing precise berry count and spatial information.

Grape cluster architecture and compactness significantly impact yield, quality, and disease susceptibility. These traits are complex, influenced by factors like berry size and arrangement, and are challenging to measure accurately. Current methods, such as visual scoring and computer vision, have limitations in precision and scalability.

published in Plant Phenomics on 27 Jun 2024, researchers aimed to use the Segment Anything Model (SAM) to segment grape berries in 2D images without additional training, thereby enhancing the accuracy and efficiency of analyzing cluster architecture and compactness for improved vineyard management and breeding programs.

The study used the SAM algorithm on a population of 387 vines and 1,935 clusters, generating 215,090 masks. For 99 vines, clusters were imaged at four angles, resulting in 3,431 images. The algorithm identified various objects, filtering out 55,550 masks of overlapping or improperly sized berries, leaving 153,939 true berry masks.

The average berry count per cluster was 44.87, with counts normally distributed. Processing time varied with grid density; a 32 x 32 grid took 55 seconds per image on a CPU and 14 seconds on a GPU. Increasing to 62 x 62 points increased processing time to four minutes and 45 seconds.

Berry counts from cluster images showed a high correlation with manual counts (R² = 0.93) but were underestimated by about 50%. This underestimation was consistent and correctable with linear regression, improving accuracy to an adjusted R² of 0.8723.

Berry size predictions were more variable but also linearly adjustable (adjusted R² = 0.8457). Imaging angle significantly impacted berry count predictions, especially for clusters with asymmetries, while berry size was less affected. The methodology demonstrated sensitivity to cluster architectural features and genetic variance, with consistent repeatability for traits like berry count and cluster compactness.

The study's senior researcher, Diaz-Garcia, says, "We emphasized the critical importance of the angle at which the cluster is imaged, noting its substantial effect on berry counts and architecture. We proposed different approaches in which berry location information facilitated the calculation of complex features related to cluster architecture and compactness.

"Finally, we discussed SAM's potential integration into currently available pipelines for image generation and processing in vineyard conditions."

In summary, this study used the SAM algorithm to accurately segment grape berries in 2D cluster images, correcting a 50% underestimation using linear regression (adjusted R² = 0.87). The findings highlight SAM's potential for precise, scalable cluster analysis in vineyard management and breeding programs.