Non-invasive smoke-taint detection in berries from grapevine (Vitis vinifera L.) using near infrared spectroscopy and machine learning models
I have been spending some whole nighters working on this topic due to recent news about extensive bush fires in Chile and in New South Wales, in January and February 2017 respectively. Both of these events have been named by the media as the biggest bush fire events in their respective histories. In Chile, the majority of the bush fires were in the central part of Chile, coincidentally to the majority of grapevine plantations. We do know that smoke taint has the biggest effect on berry contamination after veraison (7 days after on-set), which was about the timing of bushfires for a few cultivars for both Chile and NSW.
A recent report (February 2017) from the Victorian Government – Australia, has concluded that bush fires will increase severity and the window of opportunity due to climate change, specifically due to increases in temperature, increased frequency and intensity of heat waves and drought events.
So, what can we do about it?
As said before, I started to dig up some data from an ARC – Linkage project in which I worked as part of a team from The University of Adelaide. In that project we smoked artificially a number of cultivars to see the physiological effects of smoke contamination. From the canopy perspective and stomatal conductance, to be more specific, this effect can be explained through the chemical reaction between the main compounds found in smoke: carbon dioxide (CO2) and carbon monoxide (CO) and water. When getting in contact with stomata, smoke gases they can pass to the sub-stomata cavity, which is at 100% relative humidity. These compounds are then mixed with water forming carbonic acid (H2CO3) which reduced pH (acidic) hence close stomata.
This effect was reported in a poster in which I did a model to detect this effect on stomata conductance to discriminate canopies that have been contaminated or not with smoke. See posting by clicking here. The models worked really well for all the cultivars studied but Sauvignon blanc. I did attributed this effect to the morphology of leaves for this cultivar, which have high pubescence in the abaxial side. My hypothesis was that this offers a barrier to smoke, which can explain the inefficiency of the model based on the lack of stomatal conductance reduction.
I am currently working in the development of these models considering the top of canopies to apply them using infrared thermal imagery from Unmanned Aerial Vehicles (UAV), which can map a whole vineyard in the days after a bush fire event.
After the bush fires in Chile, I have revisited the Near Infrared (NIR) data from berries to see whether I was able to generate machine learning models to detect smoke taint in berries triggering the instrument around the skin and then measuring also in halved berries. The instrument that we used was a ASD FieldSpec® 3, Analytical Spectral Devices, Boulder, Colorado, USA. Which measures in the range of 350 – 1880 nm.
I have decided to concentrate the models in the700 – 1100 nm range using the second derivative of data, since it is the range of inexpensive NIR instrumentation. Also, after analysis of the rest of the wavelength range, the improvements of models obtained did not justified the jump in price of instrumentation from around $2,000 – $3000 to $38,000 – $65,000.
I did obtain three interesting models, the first was to detect whether the berries measured in a bunch have been contaminated or not using Artificial Neural Networks. I tried with data from whole and half berries and surprisingly I got better results with full berries. This is important since it renders the methodology as non-invasive. And makes sense when reading reports which found that the majority of glycocongugates are found in the skin of berries, which are higher than the pulp and higher that those found in seeds.
Then I tried models using machine learning fitting algorithms to see whether I could predict the levels of glycocongugates in the berries (whole):
And finally, whether I was able to predict smoke taint compounds in the wine made with contaminated and non contaminated berries, such as guaiacol, Syringol and cresols:
These are very exciting preliminary results which I am in the process of writing up for a peer review publication. The best thing is that if measurements made in the field are associated to GPS tagging, this could produce a contamination map using simple kriging interpolation techniques. This tool can support the decision making process towards differential harvest to avoid contaminating the whole production and salvaging fruit that has not been contaminated to the levels of spoilage. It offers also alternatives for winemaking to avoid excessive crushing and fermentation with skins that could contribute to the increase of smoke taint compounds in the final wine. Decisions can also be made by assessing the timing of ageing in barrels and whether it is required at all. And obviously, making non-contaminated wine using non-contaminated fruit. As can be seen in my preliminary results, the model for glycocongugates rendered a 10% error, which by quantifying the levels of smoke taint compounds does not make much difference in the final wine and it may contribute even to increased organoleptic characteristics with tones of leather, wood, bacon, etc
Something to look forward after all these tragedies.
Dr Sigfredo Fuentes
The University of Melbourne