International Epidemiology Workshop 2024 – 9th -12th April 2024 – Foz do Iguaçu (Brazil)
If you’re reading this title in an AgriFoRwArdS CDT blog post, the obvious question is – “What has an epidemiology conference got to do with agricultural robotics?” So to give you a bit of background, this event happens once every 4-5 years and is one of the main gatherings of plant epidemiologists from across the world. Plant epidemiologists study the spread of plant diseases and pests so it’s about protecting crops and natural ecosystems rather than preventing coronavirus. The focus of my PhD is using reinforcement learning to optimise the detection and control of invasive plant disease epidemics. I’m taking an idea used heavily in robotics (reinforcement learning) and applying it to plant epidemiological models with the aim of minimising the impact of these diseases where resources for detection and control are limited (i.e. most of the time). Landscape scale models of invasive plant disease are one of the specialities of the lab I work in – Nik Cunniffe’s Theoretical and Computational Epidemiology group in the Plant Science’s department in Cambridge. So this is a key conference for our lab and a chance for me to get wider review of my work from the epidemiological community.
For the past few months I’ve been focussed on finding a sensible optimisation method to compare my reinforcement learning results to or, to ask the question a different way, what should the baseline be for this work? You can’t compare optimised disease control to “no control” because that is always going to be better (unless your optimisation is _really_ bad) and you can’t compare to an analytical optimum in most relevant cases because if you could find an analytical optimum, then you wouldn’t need to use a method as complicated as reinforcement learning! In robotics, we have a series of commonly used benchmark environments (e.g. the MuJoCo environments in Gym/Gymnasium) and standard algorithm implementations (e.g. StableBaselines3) and the idea of benchmarking different algorithms is well established. In epidemic control, the field is more fragmented with different papers presenting different optimisation methods against different disease models, with different potential mechanisms and limitations for control and different objectives. In this scenario, it’s not so easy to just “take the method from the last good paper” even when there is code available because it’s unlikely it will be applied to a system that is exactly the same as your system and adapting the methods can involve a lot of work. To get around this, I’ve compiled a spectrum of different approaches to baselining ranging from basic heuristics up to a full implementation of previous methods that can be used to help epidemiologists think about how they can baseline their new methods for optimisation. I condensed a lot of this thinking along with an example from some recent work into a poster “Baselines for Prioritisation of Epidemic Control” (archived here) which got some good interest in the poster sessions. It also links in to some work we’ve recently submitted for review as a paper (preprint link here).
My supervisor Nik presented “Modelling fungicide resistance management strategies: progress and challenges” which summarised the fungicide resistance work that has been ongoing in the lab for several years. Fungicide resistance is when a fungal pathogen evolves so it is no longer affected by a particular fungicide. The practical effect of this is seen in the field when growers report that diseases that were previously controlled by fungicide sprays are damaging crops and so growers and agrochemical companies are all interested in methods which can slow down the rate at which pathogens become resistant or the rate at which resistance spreads. Two main strategies have been investigated in the modelling literature – alternating (where you spray fungicide A at the first application and fungicide B at the second application some time later) and mixing (where you spray both fungicides at half doses for both applications). The work Nik presented included various investigations of this idea for monogenic resistances, as well as to the case of polygenic resistance where the resistance of the pathogen to the fungicide is governed by more than one gene. If you are interested in the results and details, the main papers covered in the presentation are here, here and here.
The main benefit of attending this conference for me was to get a wider view on the range of epidemiology work carried out across different subfields and across the world. The first presentation at the conference was from Larry Madden who wrote several key textbooks used to teach the subject in the US – people were coming up to him at the conference and asking him to sign copies! He gave an overview of how plant disease epidemiology has developed in terms of both ideas and key researchers over the previous decades using the past IEW meetings as milestones. He also had a very active twitter feed over the whole conference (https://twitter.com/thelarrymadden)!
In general, because the conference was in Brazil, there were a lot of representatives from US and South American research labs. I got to meet new people from these groups as well as strengthening links with other labs that I already knew in Warwick, Reading and INRAE. Hearing from the American teams highlighted the different priorities for these labs with much more opportunity and requirement for direct grower involvement via extension schemes and a focus on weather based disease forecasting models for larger growers and coalitions who were directly funding research. On the computational side, concerns for these labs were around data management and better interfacing to growers’ planning systems – the growers understandably want one interface rather than a separate app for each different model that they are using. For the weather based forecasting models, complexity is increasing in two contrasting ways – either data driven approaches using complex statistics and machine learning or model driven approaches which look to use targeted experiments and information from literature to justify adding parameters to a mechanistic model. Concerns were raised in some of the questions about whether the ML models – using time series weather data or satellite data to inform disease prediction – would generalise or find acceptance with growers. The generalisation concern was also seen for cases where remote sensing data from drones or planes – usually multispectral or hyperspectral images – were used for early disease detection. These were very similar issues to those I see in discussions for agricultural robotics. My opinion is that, on the epidemiological side, we need to get more fluent and comfortable as a field in talking about quality metrics for ML approaches. Everyone loves a p-value but very few people were talking about train-test-validation split or evaluating how well the dataset used for training was likely to generalise wider applications e.g. predicting what is going to happen as the climate changes.
If you’ve managed to get through all that serious stuff, we also had a fun time. Foz do Iguaçu is on the border between Brazil, Argentina and Paraguay and is famous for a series of waterfalls. The organisers made the sensible choice that they didn’t want people to miss sessions to see the falls and we all went on a big coach trip together. The falls are in a forest area and in a group full of plant, fungus and insect specialists, you had the comedy situation where people were taking almost as many pictures of the insects and fungi in the forest than of the huge waterfalls…
Thank you to Rachel for taking the time to share her experiences of this fantastic sounding conference.
If you have something you want to share with the AgriFoRwArdS community, please get in touch at agriforwards.cdt@lincoln.ac.uk.
Be First to Comment