Marty-Mooij

Marty Mooij

 

 

[Image of a split circle appears with photos in each half of the circle flashing through of various CSIRO activities and the circle then morphs into the CSIRO logo]

 

[Image changes to show a new slide showing a photo of a harvester harvesting cotton and Marty Mooij can be seen inset in the top right talking and text appears: Towards an OFE App, Design and development in parallel, Marty Mooij & Ross Darnell, March 2021]

 

Marty Mooij: So Ross and I will be talking about an on farm experimentation app and how a design team and a science team together developed this app to help farmers run better on farm experiments.

 

[Image changes to show a new slide showing a photo of a crop, and Marty can be seen inset in the top right talking, and text appears on the slide: On farm experimentation – OFE, Why farmers do them, Will this new fertiliser work for me?, Some of the problems they face, I do not want it to impact my yield too much, it should not be too difficult to do]

 

So what is on farm experimentation and why do it? An on farm experiment is a way to, at small scale and with relatively low risk, learn about new ways of running things on your farm. There is, however, an issue that not all farmers are actually using on farm experimentation and some of the issues around that are that for instance, even if you run it on one or two fields there might still be a small impact on your yield and if you’ve had a couple of years in drought and you try something new, doesn’t work out, you don’t have a lot of margins in your business, so you don’t necessarily want to do that. Another thing is that they can be a bit more labour intensive and that might put some people off too.

 

[Image shows new text appearing on the right of the slide: Our initial ideas, Give them an app, Should make it easy to plan and manage OFE, Encourage them to do more OFE, Grains and cotton as good first candidates]

 

So what we kind of had as an idea was to give farmers access to an app that should make it easy to plan and manage on farm experimentation and by making it easier, encourage them to do more. So if it’s easy to do, why not do more and learn more? And as a start, we were focussing on kind of grains and cotton crops because they were good first candidates, as in those, in those sectors, they actually use, on harvesters, they actually have yield monitors so they measure precise yields while they harvest and that makes it a lot easier to measure the outcomes of these trials. So to measure, was this area where I applied this fertiliser more successful than where I applied my old fertiliser?

 

[Image changes to show a new slide showing two diagrams showing three coloured strips representing no fertiliser, old fertiliser, and new fertiliser and a fertiliser key can be seen on the left, and another diagram can be seen on the right showing variability of a paddock, and Marty can be seen inset in the top right talking, and text appears on the slide: How OFEs work, No replicates, One replicates, Managing variability]

 

How do they actually run these trials? And here is a bit of an oversimplified example, just to get the point across. There’s a little bit more subtlety and nuance involved than this but, so let’s say I wanted to try this new fertiliser. Well, I need to pick a, a representative field on my farm where I want to start trying this and then I need to kind of break down the different treatments that I want to apply and want to test across that field. And so in this case the simple, oversimplified example, is I have a strip of new fertiliser, the yellow strip, a strip of the old fertiliser, and a control strip that I run as well. And then basically, I just guide the fertiliser across, the different fertilisers across these different parallel strips, basically grow the crop and then at the end of the season, I measure which one had more yield, the new fertiliser or the old fertiliser and then I know which one might be more applicable. But just assigning kind of a single treatment to a single strip like in the top example might not always give me the best insights of what works most, what works best on that field. So if you look at the example on the map on the right-hand side it might be that the dark blue area is an area that doesn’t really drain all that well and so it’s always wet, and the orange area might actually drain better than some of the rest of the field and so it’s always a bit dryer than the rest. And so if I only use kind of those three strips as in the top one without the replicates, so if I don’t use any replicates, then you see that the orange area is only really covered by the white strip and a little bit of the blue strip. So I’m probably not going to learn anything about the new fertiliser in that area.

 

[Image changes to show a new slide showing a flow chart showing growers moving through steps to apply learnings, and Marty can be seen inset in the top right talking, and text appears on the slide: What we considered the key problem, Growers, Uncertainty – What about this new fertiliser?, Need for trial – How do I test if it works for me?, Do trial – Gather and review data, Lessons Learnt – Where else can you apply learnings, Apply learnings – Apply learnings more broadly across farm, Advisors, Suppliers, Extension, Grower groups, Provide support, Can learn from]

 

If you look from left to right, you’re going to see the whole process of running an on farm experiment. So, I’ve got a question about my fertiliser. I run the experiment all the way to collect the data, create the insights and then apply those learnings more broadly across my farm. Now if I picked a design that was sub-optimal to answer the question that I have, I produce data that basically can be inconclusive or actually give me the wrong conclusion. So having a good design for your experiment is critical to having good data which you can then provide meaningful learnings about. And so that’s where we decided to focus on helping farmers design their experiments in such a way that at least the data that comes out of them is meaningful and can provide meaningful learnings further on. And then we also just focussed on the farmers themselves because as I said at the start, the easier it is for them to do this themselves the more often they are hopefully running on farm experiments to learn about, about new and better way of doing things. So our focus within this project was on trial design so experiment design for farmers.

 

[Image changes to show a new slide showing a table explaining the workings, and Marty can be seen inset in the top right talking, and text appears on the slide: The way we worked, Problem to solve, Develop concept, Test concept – How can we help farmers choose an OFE design that helps them answer questions they have, How can we make statistical significance insightful for farmers?, Workshops and interviews, Science – Make sure that OFEs have statistical significance, Minimum detectable differences given an OFE design layout, Minimum detectable difference centred around average yield for that paddock]

 

So how did we go about doing this? And this is the bit where I was talking about design and science working very closely together. So if the initial question was, “How can we help farmers to create designs, experimental designs that help answer their questions?”, then the science team that we’re working with were like “Well, we need to explain concepts like statistical significance to farmers”. And then like I’ve had some stats courses in uni. and I already have issues understanding statistical significance. So the question then became, “Well, how can we make statistical significance insightful for farmers?”. Then we started talking about, “Well, it’s actually about minimum detectable differences”. So if you measure a difference of yield between the old fertiliser and the new fertiliser of, for instance, one bale of cotton per hectare, is that significant or not? And in some cases it might be and in some cases it might not be but that’s kind of the language that we wanted to use because most farmers understand from year to year the yield varies, from field to field the yield varies. So, that’s terminology that at least they can understand and can use to reason about whether or not the outcome of the experiment was what they were expecting or not.

 

And so once we landed on this minimum detectable differences – and I’ll show you in a second what that looks like – we had a number of workshops and interviews with farmers and advisors to test those ideas and then based on that we built on that. So what does that look like, that minimum detectable difference concept?

 

[Image changes to show a new slide showing a diagram on the right of strips of fertiliser on a field, and information about the application of the fertiliser on the Coleambally Cotton Farm appears on the left of the diagram, and Marty can be seen inset in the top right talking, and text appears on the slide: How we tested, Test participants, 5 growers, 5 advisors]

 

So down the bottom, you can see that or you can basically change, at the top you can change the number of treatments and replicates and then down the bottom you will see what the impact of that is so if you end up in the grey bar with those treatments and replicates you see that, basically, it’s not statistically significant, so you can’t say if it’s chance or actually the experimental treatment. If it ends up in green then you can say that the treatment led to better results and if it ends up in red, then basically it led to worse results. Tested this and we had a number of farmers say, “Great, we want to do more and we want to analyse our own experiments using this way”. And this is why we built a prototype that Ross, in one minute or more, is going to talk about.

 

[Image changes to show a new slide showing an inset of a website showing a field view on the right, and Ross can be seen inset in the top right talking, and text appears on the slide: Read in historical data, Required for MDD calculations specific for that field, Read in GIS shape/raster file of historic yield map, Select variable of interest, Field boundary detected, Start row detected, Field dimension calculated]

 

Ross Darnell: Thanks Marty. So to get a, to design the experiment and to calculate this minimal detectable difference, we needed to get enough knowledge about the amount of variability there is in the particular field that we are looking at. So with this, we use historical yield maps to do that. So we ask the user to upload a historical yield map.

 

[Image changes to show a new slide showing an inset from a web page showing strips in a field, a graph showing minimal detectable difference, and information about the application of fertiliser, and Ross can be seen inset in the top right talking, and text appears on the slide: Provide possible designs, Whole of block strip plots, Input harvester width, Ask for number of treatments (+ labels), Possible designs suggested given field dimensions, MDD calculated for different possible designs, User chooses design and output to script file]

 

We’re looking at whole of block strip plots, we do have an updated version that does other types. We asked the farmer for the width of their harvester. We asked them for their treatment numbers and then we provide them with the possible designs and the comparison in the figure at the right top there gives an idea about the different types of minimal detectable differences they would be likely to get given the variability they have observed in their previous yield crops and then it represents that. It gives, it outputs that experiment to a particular, script file for the farmer to use and to implement.

 

[Image changes to show a new slide showing to show an inset of Analysis of Experimental Design data from a web page, and Ross can be seen inset in the top right talking, and text appears on the slide: After the OFE is completed, Already have design parameters, Upload new response file (NDVI/yield), Select response variable, Output results plots/tables]

 

Now the, after the experiment is done, the information about the design is already there. The farmer needs to upload the next field map that’s from the response and then they can provide, then we provide some sort of measure of comparing the results to different treatments across the farm. And there are various, we’re working on various ways of outputting that result in plots and tables of various forms.

 

[Image changes to show a new slide and Marty can be seen inset in the top right talking and text appears on the slide: What is next?, Engaged with some RDC’s (Grain and Cotton) as they are already engaged in OFEs and can deliver the app at scale to their farmers, They showed an interest in these tools, Conversations with RDC’s regarding uptakes is yet to happen]

 

And what’s next? We’re engaging with some RDCs, and they are already engaged with, in the OFES with scales to some farmers, to some farmers that showed interest in these tools and we’re having conversations with the RDCs regarding uptake. Thank you.

 

[Image changes to show a white screen and the CSIRO logo and text appears: CSIRO, Australia’s National Science Agency]