[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 map of Australia on the main screen and Randall can be seen inset in the top right and text appears: Space-based monitoring of Australian paddock-scale crop yields, Randall Donohue, Roger Lawes]
[Image changes to show a full size Australian map showing the wheat production areas and text appears: Australian grain product statistics, Wheat production, 2016 agricultural census, Australian Bureau of Statistics]
Randall Donohue: This map that you can see is Australian Bureau of Statistics wheat production census data and a few years ago this was, that was, about the best information you could get if you wanted to know how much food was being grown and where and what type. So you can see that it’s no more accurate than a regional average. It’s usually produced six to twelve months after the event. So it’s not, I mean it’s the best we’ve got, but we can do better.
[Image changes to show a new slide showing the C-Crop logo and Randall can be seen inset talking to the camera and text appears: Graincast, A three-fold approach – Linking satellite and farm-based sensors, Satellite-based mapping of crop type, Satellite-based prediction of crop yield, Graincast, the digital infrastructure that maps, monitors, manages and forecasts the Australian grain industry’s production across vast areas of our continent]
So Graincast is a project which I was involved, it had a three-fold approach. One was to figure out how to link the satellite streams of data that we have to on-field information. So you’ve got to have field-based data to, to develop and calibrate models so that part of the project was about getting an app designed and put into the hands of farmers and then there’s a two-way exchange of information, and they can tell us about what they’re growing, we can give them predictions about what’s going to happen. That’s the first aspect, the second aspect was, we have to know what’s grown there, so crop type mapping was the second part. So, as the season goes on, how can we use satellites to predict where, which fields are wheat, which fields are canola, barley, lucerne etc. The third aspect was, once we know what’s growing there, can we predict how much it will yield? And that’s where the work that I was involved in came into and that’s why we developed this model called C-Crop which is a satellite-based crop yield prediction model.
[Image changes to show a new slide showing a diagram of a C-Crop model and Randall can be seen in the top right talking to the camera and text appears: C-Crop, Is a national, paddock-scale, satellite-based crop yield model, Designed to be as simple as possible, Is fully scalable, Sits within CSIRO’s Digiscape platform]
The, the very general characteristics of the model is it’s national so to apply it more nationally, it has to be very simple and not require a lot of data, specific data as inputs, but paddock scale so it still has to be able to operate at a very fine scale, accurately. It’s fully scalable so you can use it to derive sub-paddock information if you need to, regional, state, national, it’s the scaling and it currently sits on Digiscape’s platform and is run operationally at CSIRO now.
The general bare bones of the model is, it’s a very simple carbon turnover model so we use the satellite data to predict the photosynthesis of a crop, so how much carbon is being captured by the crop initially. Some of that carbon gets put into the roots and we don’t care about it much after that, some of the rest goes above ground. The plant burns an amount of carbon just to keep alive and grow, so there’s a loss of carbon through respiration. What’s leftover is called net photosynthesis and that’s the amount of carbon that the plant has available to grow, build biomass. And once we get to the end of the season, there’s a certain amount of plant biomass and from that we predict grain yields, so that’s the sort of the mechanics of the model.
[Image changes to show a new slide showing two line graphs for the model and text appears: Model details, Inputs – crop type, elevation, air temperature, satellite greenness, 250m resolution, 16-day time step, Canola, wheat, (barley)]
For a national model, it’s simple, so it only needs four inputs, data inputs, which are quite generic, crop type is the hardest one, so that’s why the previous part of the project was so important. Elevation, air temperature and a measure of greenness from the satellites so how much green leaf coverage is on the ground. The model runs at 250m resolution, and has a 16 date time step. It has fixed sowing and harvest dates and we found that that didn’t effect model performance even though they’re quite unrealistic in some places. We’ve used it to predict canola, wheat, reasonably well, barley not so well.
And the way the model works so, on the bottom right, start the model going at the sowing date. We track the photosynthesis and then we model how much of that photosynthesis, photosynthetic carbon gets put into the crop mass that goes on. Throughout the growing season the crop grows, we have a certain amount of mass at harvest date and then we, we associate the mass at harvest to the final grain yield. That’s the method.
[Image changes to show a new slide showing a Header yield map on the left, and then two Australian maps showing where canola and wheat is grown and a text heading appears: Calibration]
We calibrate the model using these yield maps, they’re pretty cool. On the left, it’s 5m resolution paddock yield data obtained from tractors as they run past the field. These were critical actually to the, to the project being able to go ahead so that was a major step forward just having that, that data available.
[Image changes to show a new slide showing two graphs on the left showing paddock-level accuracy and then a blue coloured patchy block and a green coloured patchy block and text appears: Performance, Paddock-level accuracy, Paddock-scale estimates are currently about 70% accurate]
And this is what it looks like at a paddock level, so canola and wheat. We’re predicting with about a 30% error. So that’s not quite as good as what we would like, 10% would be excellent. But for a very simple model that requires no specific ground information to run, apart from crop type, that’s pretty good so we’ve proven that this one, one continuing focus. On the right the maps, just the blue, the blue map shows you the information we’ve got about what’s growing there, so that’s the crop type and then you overlay that with how much is growing there and you’ve got your yield production information system.
[Image changes to show a new slide showing a yield forecast graph using C-Crop and Randall can be seen in the top right talking to the camera and text appears: Forecasting with C-Crop, C-Crop is calibrated at various stages in growing season to predict final yield, A yield forecast is made every 16 days from early June onwards]
We also looked at how you can use this in forecasting mode, so for every time step during the growing season from June onwards, we take the biomass at that point in time and see how well we can predict the end of set from harvest forward from there.
[Image changes to show a new slide showing a bar graph showing Forecasting Grain Yield and then two line graphs appear on the right and text appears: Paddock-scale forecasting, Reasonable yield forecasts can be made from August onwards]
What we’ve found is that from about August onwards the, the forecast or predictions are reasonably good.
[Image changes to show a new slide showing four line graphs showing yield forecasts for WA for wheat and grains for 2018 and text appears: State-scale forecasting, In WA for 2018, CSIRO estimated the grain yield to within 0.4Mt, CSIRO yield 16.36 Mt, GIWA yield 16.76 Mt]
Prior to that, not so good and that’s because the crop really hasn’t developed yet enough for the satellite to be able to tell you much of use about what’s grown there and then if you scale up to a state level, the accuracies of the predictions get a lot better. So this is just an example for WA for 2018. So the yield predictions are pretty good, the production which is, when you times the yield by the area, you get your total tonnage. It’s the area that’s a bit trickier to get right actually, we’ve found. But anyway the, as you scale up to a region or a state level, errors start cancelling out if your model is unbiased and your errors plummet so this is a pretty, pretty accurate forecast that we’re able to do.
[Image changes to show a new slide showing a map of South Australia, a map of Australia with South Australia circled, an inset part of the map enlarged, and then enlarged again and text appears: A new generation, 2 years ago, Now]
So this is, this is where things are at now. So two years ago, on the left there, that was our best estimate of national yield, grain yield. On the right is what we can now do, which is nationwide, but paddock scale so you can see on the right you can zoom right down into an individual paddock, know what’s growing there, how much we yielded.
[Image changes to show a new slide showing Randall in the top right and text appears: Thank you for your attention]
[Image changes to show the CSIRO logo and text appears: CSIRO, Australia’s National Science Agency]