Artificial intelligence and machine learning to improve Australia’s winemaking industry

Winemaking contributes over A$ 40 billion to the Australian economy each year. Among the many challenges being faced by this centuries-old industry are managing pests and diseases, producing a consistent crop and using water efficiently.

According to the report made by the University of Melbourne, artificial intelligence and machine learning have the potential to change the face of winemaking.

Dr Sigfredo Fuentes is a plant physiologist and agronomist at the University of Melbourne. He explained how they are able to make use of cutting-edge technology to keep wine racks stocked with high-quality drinks.

Drones improve the irrigation management of wineries by taking detailed pictures as they fly overhead. Multi-spectral and thermal infrared cameras that are mounted on drones can pick up signs on the vines that indicate their water status.

The machine-learning model to assess stress was developed using ten vegetation indices as inputs and is able to classify the plants into three levels of stress: absent, moderate or severe. With an 83% accuracy, the winemakers are able to use water and fertiliser supplies more efficiently.

Moreover, drones can also provide information on which parts of the vineyard are hit by disease or pests, as well as what plants have died and need replacing. This proves how technology allows a 45-hectare land to be surveyed in 15 minutes and have the data ready a day later.

Machine learning algorithm has aided in grapevine cultivar classification, assisting on telling grape varieties apart. It can also give information about water stress and fertiliser status.

Dr Fuentes explained that their machine learning model uses 13 morpho-colorimetric parameters, which are shape and colour measurements, as inputs and predicts the cultivar with a very high accuracy of 94%, and water stress with an accuracy of 88%.

The algorithm can also be a part of a computer application, which can be downloaded to a smartphone or tablet PC.

Grapes that have been in close proximity to a bushfire can produce smoke-tainted wine, which has a distinct, unpleasant smoky and leathery taste. But using the naked eye alone, to determine which grapes have been contaminated, proved to be difficult as they look the same as the uncontaminated ones.

Drones have aided the team of Dr Fuentes in addressing the issue. They have developed a way to use thermal infrared imagery to measure the pattern of temperatures in the canopies of the vines. Thermographs are images in which every pixel is a measure of temperature rather than light.

Dr Fuentes discussed that smoke contamination disrupts the vines’ temperature, so measuring this thermal pattern and analysing it through machine learning models allows them to determine which plants have actually been affected by a bushfire.

Their algorithm is able to provide an accurate map of what areas of the vineyard have been affected by smoke, thereby providing growers what they needed to make informed decisions when harvesting their grapes to avoid contamination.

In order to determine whether a grape is ready to become quality wine, its sugar content should be assessed. More important to assess though is the production of flavour and aromas, which according to research is related to the pattern of cell death in grapes.

He explained that a percentage of the fruit must be dead in order to produce the different aromas and flavours for winemaking. To achieve this, they made use of a handheld device that uses near-infrared wavelengths to measure the level of grapes and the patterns of cell death, made possible by machine learning algorithm. This allows winemakers to see in real time which grapes are ready for picking.

Estimating yield early in the season is particularly important to help winemakers plan logistics including allocation of resources, like water and fertiliser, how many staff to hire for the season, and how many barrels to have ready, among others.

Dr Fuentes shared that big data and machine learning are used to predict seasonal yield faster with an accuracy of around 80 – 90%. They take the historic data of a specific vineyard such as soil data, management data, meteorological information and actual yields per season.

They would then plug the data into their machine learning models to predict the yield from the coming season, from early stages of growth thereby aiding winemakers plan for that season much more effectively.