On Thursday 12 December 2019 at AgroParisTech Paris, Mathilde Chen defended her thesis, jointly supervised by François Brun (ACTA) and David Makowski (INRA). Her three years of research on the dynamics of the appearance of mildew in the Bordeaux vineyards and the use of fungicides have generated new knowledge and raised new questions. Summary and prospects.
What were the context and the objectives of your research?
The goal of the EcoPhyto plan is to reduce the crop protection products used in France by at least 50 %. My thesis falls within this context. In viticulture, these products are still very present, in particular the application of fungicide products against grapevine downy mildew. In the Bordeaux region, my area of study, anti-mildew treatments account for 43 % of all crop protection applications (averages for 2010 and 2013).
Today in viticulture, it is common to treat vines before the appearance of the first symptoms of mildew and this strategy prevents us from knowing the actual date of their appearance. The goal of my research was to model the dynamics of the appearance of mildew. The aim was to represent the evolution of the proportion of plots affected by the fungus, then to compare these dynamics with the dates of first application.
Which data sources did you use?
I had several information sources on the date of appearance of disease symptoms and the climate for the plots studied:
- A regional IFV database includes data from epidemiological observations in a network of untreated controls (untreated sections of plots). Every week, from 2010 to 2017, engineers from IFV and its partners noted the number of diseased plants, leaves and bunches. This data was used to estimate the real date of appearance of the first symptoms and to represent the dynamics of disease emergence at the regional level.
- A weekly forecast of the date of appearance of the first symptoms was also produced by local experts, viticultural consultants, researchers and winegrowers, in plots that they monitored between 2017 and 2019. They used the MATCH Elicitation Tool® to formalise their estimation of probable dates, in the form of a probability distribution. This expert estimate can be combined with results from the analysis of epidemiological data.
- Finally, climate data were taken from the Météo-France database to which I had access.
Which methods and models have you used and developed?
The epidemiological dataset was incomplete, so at the beginning of my thesis I used survival analysis techniques to analyse it. These are generally used in medicine, for example to calculate the time between the start of medication and patient recovery. In my case, the patient was the vineyard plot! In the first stage, I focused on the time between the beginning of the year and the date of appearance of the first symptoms. My results revealed interannual variability in epidemics. One key finding: 29 % of untreated controls in our network did not present any mildew symptoms at the end of the season (on average, between 2010 and 2017). However, although the symptoms appeared at an earlier or later date, the crop protection practices did not necessarily change from one year to another.
Does using the date of appearance of the first symptoms of mildew to determine the first fungicide application help to reduce fungicide use? Can this date be used as an indicator of the health of plots at the end of the season? I addressed these questions in the second part of my thesis.
To estimate the impact of the date of first application on the number of applications, we calculated their number in two cases: when the first application takes place on the date of appearance of the first symptoms in the plot, and when the first application is determined according to current practices in the Bordeaux vineyards. Finding: postponing the first application to the date of appearance of the first symptoms makes it possible to more than halve the number of applications in relation to current practices.
I also used machine learning methods to predict the risk of reaching very high levels of mildew contamination at the end of the season. I observed the health of plots at the end of the season based on the regional dataset, and I classified plots according to their level of contamination as “severely infected” or “mildly infected”. It was necessary to adapt several classification algorithms to this case (Random Forest, Gradient Boosting, LASSO type penalised regression methods). These prediction methods were used to predict – based on the date of appearance, weather data or the two types of explanatory variables – whether the risk of mildew would be high or low, by plot.
Early onset of the mildew epidemic is strongly linked to its severity. The statistical models developed show that the earlier the symptoms of the disease appear, the higher the probability that mildew will be severe at the end of the season. This may be linked to the fact that the Plasmopara viticola fungus responsible for grapevine downy mildew has several cycles in one season, and the longer it has to complete these cycles, the more the disease will spread.
We can use these predictive statistical models to determine treatment in the case of high risk of mildew. With, as a rule to decide on treatment: “based on the observed date of appearance of the first symptoms > prediction of risk at the end of the season > if the risk is very high, we treat, if not, we wait”.
What are the prospects?
This research provides new knowledge, shows the evolutions possible and raises other questions. New studies assessing the impact postponing the date of first application has on control of the disease will be needed before considering their integration in a decision support tool. One option could be to offer farmers crop insurance based on better use of applications. Insurance companies are already working on this in the Bordeaux vineyards, and these models could contribute to it.
And now, what are your post-thesis plans?
I am about to begin a postdoctoral position at Inserm, in a Parisian research unit specialising in human ageing. This is still epidemiology, but I am leaving plants to focus on the use of data analysis methods to estimate the impact of certain factors on the development of diseases linked to human ageing, in particular, cardiometabolic diseases such as diabetes, and diseases linked to dementia.
In the future, returning to the field of agriculture to study the impact of agricultural practices on human health is a research subject that would interest me greatly!