Winegrape yield forecasting tool based on interpretation of video data using machine learning methods.
Annual yield variability in wine grape vineyards impacts most aspects of the supply chain but is particularly important for decision making at the winery. Current methods commonly have an error of around 30%, are labour intensive, and only assess a tiny fraction of the vineyard block. CSIRO, with the support of Wine Australia, has developed tools to provide large-scale estimates of yield potential post-budburst, yield prediction post-fruit set and yield estimation pre-harvest.
We are seeking expressions of interest from individuals and/or organisations who are willing to engage in the commercial development of the system. Development of the technology should have a primary focus on providing impact for the Australian viticulture industry but can have application globally, as well as be used in other crops.
Partnership options include, but are not limited to:
- Commercial license of the technology in its current form.
- Co-development of a commercial hardware product and/or data processing for integration with an individual's and/or organisation's own algorithms and systems.
- Co-development of a commercial hardware product with CSIRO, including novel algorithms for a unique combined product/service.
Depending on the arrangement, Wine Australia may provide support to assist with the commercial development of the technology.
Each of the tools used in the winegrape forecasting system take video from a common commercially available device (e.g. GoPro, smartphone) mounted to a vineyard vehicle and driven along rows at moderate speed during daylight hours. The video is post-processed, using pipelines based on machine learning tools, and combined with historical information to generate a yield prediction at various timepoints during the season. Prior to harvest a direct estimate of yield is produced.
There are a number of significant advantages to the CSIRO-developed wine grape forecasting system, including:
- A large number of datasets collected across a range of growing strategies, winegrape cultivars, plant growth stages and environmental conditions with varying hardware.
- Matching physical ground truth data at a per panel or better resolution.
- A range of digital ground truth labelling strategies.
- Strategies to cope with partial occlusion of fruit, typical in ‘sprawl’ vineyards that form the majority of Australian viticulture.
- Strategies to obtain 3D position of bunches in the vineyard to minimise estimation error (e.g. due to double counting).
- Pipelines for data processing.
- A multi-season experience in deploying varying devices and collecting ground truth data under commercial field conditions.
The winegrape forecasting tools can be used in almost any Australian winegrape vineyard to generate predictions of fruit yield at various times during the growing season. The tools should also work for table grapes, though no validation in table grapes has occurred to date.
Data sets and analysis pipelines are jointly owned by CSIRO and Wine Australia. The pipelines make use of various open source software libraries.
The work to date has been undertaken by a team led by Agriculture and Food researchers, with input from researchers in Data 61 and Manufacturing.
The Agriculture and Food team has 15 years' experience in providing strategies for improved crop management in winegrapes.
The Data 61 team has a track record in developing intelligent end-to-end imaging and computer vision solutions and via collaboration with industry partners bring them to the market. Through licensing, the algorithms developed by the team have been incorporated into commercial instruments and systems. About 2,000 licenses have been sold so far, giving research labs around the world access to the most advanced image analysis algorithms.