Randall Donohue Frost Monitoring
[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 and Randall Donohue can be seen inset in the top right talking and text appears on the slide: Satellite-based monitoring of frost occurrence across Australia, Randall Donohue]
Randall Donohue: We, a whole team of us, got some seed funding through Digiscape to look at, can we use satellites to detect frost across Australia?
[Image changes to show a new slide showing an enlarged photo of a frost particle and Randall can be seen inset in the right talking and text appears: Frosts, food and satellites, Developing a national frost monitoring system, The challenge – to use satellite imagery to rapidly detect when and where frost occurs, across all of Australia, Specifications – detection of frost location, timing, duration and severity, within 12 hours of occurrence, at sub-paddock resolutions, continentally, Potential applications – Post-frost decision-point information for grain growers, Regional intelligence on likely impact of frosts on grain production, Paddock-scale frost history mapping for paddock zoning, Training next generation frost forecasting models]
So I’m going to talk a little bit about this. The challenge we set ourselves is to figure out how we can use some novel approach to grabbing satellite data and coming up with a monitoring system for frost across Australia. What we wanted to be able to achieve, eventually, is a national system, so Australia-wide, paddock-scale, continuous monitoring that can report on a frost event within 12 hours of it happening. That’s, that’s the long-term aim. This is useful, this would be useful because no such thing exists. It’s very hard to get frost data across Australia at the moment. It’s important for grain-growers, who after a severe frost event they have to make some pretty major management decisions about what they’re going to do with their crop depending on how severe they think that frost was. The grain industry as a whole can use this sort of information to predict for logistics on what eventual cross… crop yields are likely to be. As we build up a database of frost occurrence we can use it to plan and zone land for where to plant crops, where not to plant crops, crop types, as well as building a database for future frost prediction systems.
[Image changes to show a new slide and Randall can be seen inset in the top right talking and text appears: Frosts, food and satellite, Developing a national frost monitoring system, Our approach, 1. Derive land surface temperature (LST) from weather satellite imagery, 2. Downscale LST to 30m resolution, Scientific challenges, Detecting clouds at night, Suitable accuracy (<1°C error), Simplifying complex land surface physics into a practical downscaling model]
So there were two main challenges that we, we had to address in this. The first one was, how do you take satellite-based data at night and detect land surface temperature. So, we took a weather satellite, which is, detects every ten minutes, at a fairly coarse resolution, 2km resolution but it sits over Australia and is continuously monitoring. The challenge there was to figure out how you can detect clouds at night because when it’s dark and clouds are cold and frosts are cold, they end up looking like the same thing. So we had to come up with a new way of doing cloud detection at night and because the original land surface temperature imagery from the satellite is quite coarse-scale, it’s 2km. That’s not paddock resolution. The second challenge was to come up with a method to downscale that land-surface temperature imagery to something around 30m. So, they were the two challenges.
[Image changes to show a new slide showing a photo of a satellite in space, the world globe, and a graph showing land surface temperature during a day, and Randall can be seen inset in the top right talking and text appears: Himawari land surface temperature image, Continental coverage… every 10 minutes]
So, this is the imagery. So we’re using Japan’s Himawari weather satellite. You can see on that bottom left there that it captures Australia quite easily in one image and it’s doing that every ten minutes for 24 hours a day and the temporal dynamics in this data are wonderful.
[Image changes to show a new slide showing an image of the coast of South Australia, and a graph showing air temperature, and Randall can be seen inset in the top right talking, and text appears: Accuracy of LST at 2km resolution, Currently has an error of about 2°C, Note – air temperature is not the same thing as LST]
So the satellite doesn’t actually detect land-surface temperature. It detects something slightly different. We have to convert it to that. So that was the first step and we found out that we can do that fairly well but within an error of about 2°C at the moment. There’s a bias in there, too. But the tricky thing is, is the measurements we have are air temperature but what we we’re measuring is land surface temperatures and they’re not the same things and there are no measurements for land surface temperature really, that we can calibrate to. So that’s, that’s a bit of a mismatch but this is the best we can do at the moment. And that’s not a bad effort for a quick approach.
[Image changes to show a new slide showing a satellite map of South Australia, and then a map of South Australia showing detected cloud patches, and Randall can be seen inset in the top right talking, and text appears: Detecting cloud at night, Clouds are difficult to detect at night as it is dark and they can have similar LSTs as frosted ground, Our detection algorithm is 90% accurate when it says a location is cloud-free, but only 50% accurate when it says there is a cloud]
In terms of the cloud masking, we came up with a method. It was a quick project, so it’s all about proof-of-concept rather than getting it right, and getting it accurate. We’ve come up with a good method and we know how to improve it. But on the right you can see the blockiness of the cloud masking algorithm in grey. We’ve found that it’s very accurate when it’s clear. So when the cloud masking says there’s no cloud, 90% of the time it’s correct. But when the cloud masking algorithm says there is a cloud, it’s only correct about 50% of the time.
[Image changes to show a new slide showing an animation of the land surface temperature on a map of the coast of South Australia and Victoria, and Randall can be seen inset in the top right talking, and text appears: One night of 10-min, 2km LST data, (it’s a cloudy night!)]
And this is an animation of what the land surface temperature looks like over one night. So this is just coming into sunset now. All the white squares are where clouds have been masked out. But as the night goes on, it’s getting darker. The extremely dark blue cells that pop in and out are actually probably cloud that’s not being masked out properly, especially in Tasmania there. You see there’s a lot of detail that comes from the actual land surface in the imagery that you would not get if you were using a network of BOM stations to interpolate land surface temperature.
[Image changes to show a new slide showing two diagrams showing the land surface temperature at sunset and pre-dawn, and Randall can be seen inset in the top right talking, and text appears: Downscaling LST to 30m, A windy night (fully mixed), Assumes clear-sky conditions]
So we’ve got 2km land surface temperature. Now we need to figure out how to downscale it to 30m and we had some pretty clever people doing this, it wasn’t me. And we predicted that the downscaling, it takes time since sunset, elevation, and wind speed is the three predictors or the three inputs to downscaling. So what it does is, if you take a really windy night as our first example with flat topography. A windy night means the whole near-surface atmosphere is well-mixed and so the only difference in temperature is elevation based lapse-rate.
[Image shows a wavy line appearing on the bottom of the diagrams labelled “30m topography”]
When we introduce topography into that, it’s the same process for a well-mixed windy night except some parts of the landscape are higher than others and that gives us our downscaling purely on a lapse-rate based process. So the tops of the hills there are cooler than the bottom of the valleys on a well-mixed night.
[Image changes to show a new slide showing four diagrams showing the downscaling of air temperature on a still night for sunset, midnight, early evening, and pre-dawn, and Randall can be seen inset in the top right talking, and text appears: Downscaling LST to 30m, A still night (stratification), Assumes clear-sky conditions]
The real challenge though, because you don’t get frosts on well-mixed nights. You get frost when you get still nights and there’s no mixing and more stratification so it’s a bit more complex. So this one, if you start at sunset on the top left, there you get our lapse-rate, it’s well-mixed we just assume that’s the starting conditions at sunset. And then as time goes on through the night, we have a process that introduces an inversion which is the dotted line that goes through the bottom there, the bottom of the graph on the top-right. An inversion is where air-temperature gets warmer as we go up. OK, so in this one, stratification is starting to get a little bit influenced by topography because the land-surface cools radiation quite rapidly. You can see in the bottom left there about midnight this inversion is rising through the atmosphere and below the inversion things are actually getting colder and that usually happens in the bottom of valleys as you intersect that with topography. And then at pre-dawn you can see that the bottom of the valley there is quite cold compared to the tops of the hills and so that is our process. It’s very simple. It had to be to get it done quickly and to make sure we don’t break computers doing it.
[Image changes to show a new slide showing two satellite pictures on the left of the screen, and then a graph on the right of the screen showing the air temperature over a night, and Randall can be seen inset in the top right talking, and text appears: Downscaling LST to 30m]
And this is what it looks like. On the left, there’s the 2km land-surface temperature, top, and then bottom it’s downscaled from the same image. And then on the right the dots are observed land surface air temperature, and then the dots on the line, the dashes on the lines are it downscaled. So what the graph on the right pretty much says is we’ve been able to downscale it from 2km to 30m without too much loss in accuracy. That’s a pretty good step for now.
[Image changes to show a new slide showing an animation of two diagrams showing the original imagery on the left, and then the downscaled version on the right, and a colour key can be seen between the two diagrams, and Randall can be seen inset in the top right talking, and text appears: Downscaling LST to 30m, One night of down-scaled LST, White blocks are where clouds have been detected and masked out]
Then this, this is an animation of one night again from the original imagery on the left and then downscaled on the right. So you can see that a lot of the topographic effects have come in. You can see lots of valley bottoms are much colder. In fact you’ll see a few pop up as quite blue and that’s well below zero so good frost hollows happening there, you come to see. The cloud-masking is those white dots, blobs that come in and out.
[Image changes to show a new slide, and Randall can be seen inset in the top right talking, and text appears: Next steps, Developing a national frost monitoring system, Accuracy of 2km LST imagery, More air and land surface temperature observation plus frost event observations, Improve air temperature-to-LST conversion function; achieve <1°C LST accuracy, Accuracy of nocturnal cloud detection, Higher resolution masking, better detection accuracy, Accuracy of downscaling method, Gradually increase sophistication of method to better incorporate physical processes (better energy balance modelling), Operationalise, Generate industry-specific information products, Build an automated monitoring system delivering daily reports (LST and frosts)]
So that’s what we achieved pretty much. We’ve got our land-surface temperature. We’ve got our new cloud masking algorithm. We’ve got our algorithm for downscaling. There are lots of place-holders where we want to come back and fix it up but it’s functional at the moment so we’re pretty happy with that. The aim, ultimately, is to get this into an operational system that just runs and spits out answers by about midday the next day. But one of the cool things about this is because it’s got a land-surface temperature system underneath, we can look at a lot more than just frost as well. We can look at extreme heat events during the day, for example, and the effects that that might have on agriculture, or health, or transport, or anything. But our focus at the moment firstly is getting this into some good grading prediction models.
[Image changes to show a new slide showing a photo of frost on a wheat crop on the right, and Randall can be seen inset in the top right talking, and text appears: Project team, Ian Harman (O&A), John Finnigan (O&A), Jen Austin (L&W), John Gallant (L&W), Yi Qin (O&A), Tim McVicar (L&W), Lingtao Li (L&W), Contact Randall Donohue, 02 6246 5803, firstname.lastname@example.org, Thank you for your attention]
And that is the team that did it and I thank you for your listening.
[Image changes to show a white screen and the CSIRO logo and text appears: CSIRO, Australia’s National Science Agency]