Berry Size: A Machine Vision Perspective

By Omeed Mirbod and Stephen Nuske, Ph.D.

Up until now growers in the viticulture industry have had a very sparse view of their field making it difficult to manage against variability. The aim of the sensing team is to use automated technology to give growers a vine-to-vine resolution of their field allowing them to be much more efficient in their management and more productive in term of optimizing their fields for yield and fruit quality. One characteristic in vineyards that can now be measured at high resolution is berry size using Carnegie Mellon University’s high resolution imaging units that can size the berries with submillimeter accuracy.

Berry size is a key factor which affects yield and quality and with this information growers can be more effective with precision management techniques. For example precision irrigation to manipulate the size of berries in different areas of their field to give precise control of quality characteristics like skin to juice ratio which drives acidity and tannins. As the current technology also allows growers to view the resulting berry size of their entire vineyard at high resolution, precision harvesting based on this berry size map means growers can selectively harvest based on this important quality characteristic.

Measurement of berry size for an entire vineyard can also give some indication of yield. The more growers can understand how the density of grapes changes throughout the growing season, the more accurately they can map this density into yield using knowledge of berry size. Growers can also gain insight into the current state of their vineyard. For example, in an earlier post by Cornell researcher Dr. Cain Hickey a brief case study on Concord grapes was shown that small vines tend to have small, inefficient canopies which cannot produce as much fruit as larger vines. Berry size for an entire Merlot vineyard was measured using images and indeed it was found that early season NDVI response of vines representing canopy vigor correlated to varying regions of berry size(fig. 1).  This correlation in the Merlot field was noticeable due to the high resolution stereo cameras that can size the berries with submillimeter accuracy.

Fig. 1 Berry size distribution of a 15 acre Merlot vineyard was imaged in June 2014(left) and compared to canopy in April 2014(right). Vines with low vigor in south-east region produced smaller fruit, vines with more vigor produced larger berries.

The hardware behind this automated technology is a fully rugged waterproof and dustproof stereo imaging system that is able to capture high resolution images of every vine in a vineyard. Equipped with industrial flashes that can overpower the sun’s brightness, the system can be used to collect data during the day or night. With a 5Hz imaging speed, the camera can be mounted onto an ATV or tractor and driven at 4-5 mph between rows of vines. The key however are the computer vision algorithms that take captured images and extract the information growers are interested in.

One challenge in berry sizing is adapting the algorithms to the varying texture of grapes. This change in texture can occur across multiple grape varieties or of the same grape variety but over time in the growing season. As seen in fig. 2a-b, some grapes, when light is projected on them, can reflect a shiny texture like Petite Syrah. Others, like Merlot and Cabernet Sauvignon, reflect a matte texture, while a mature table grape variety can have no consistent pattern at all. Different algorithms have been developed to accurately and with fast execution time measure berry size depending on the grape variety. More complex filters however, as shown in fig. 2c, are also under development to minimize user interaction with the software and better streamline the berry sizing process for multiple grape types.

Fig. 2 (top to bottom): a) Original image. b) basic grape texture under illumination c) advanced filtering to make grape texture look more uniform across varieties

Fig. 3 (Left to Right) Automated diameter estimation of Cabernet Sauvignon, Petite Syrah, and Flame Seedless

Fig. 5 Hand measured pixel berry diameters compared to their corresponding image detected pixel berry diameters. Two different diameter estimation algorithms were used to estimate diameter of Petite Syrah variety and Cabernet Sauvignon.

Fig. 4 Hand measured metric berry diameters compared to image detected berry diameters of Flame Seedless grape variety.

Leave a Reply

Your email address will not be published. Required fields are marked *