Precision Ag

I loaded up a new version of the MSF N Tool (1.6.1) today. It has an improved internal model and incorporates expected Spring temperature. The temperature adjustment is at the bottom of the 'GSR' tab, and is the 'anomaly' from the 1960-1990 reference period used in climate change studies. For example, putting in '+1' is 1 degree warmer in September/October than the 1960-1990 period.

Unless you modify the temperature on purpose, the tool calculates it from the years selected on the GSR tab. For the moment there's a bug in the calculations; in theory if you've selected 1960-1990, it should read 0! 

Apart from that, it should work.

I've done some more work to improve the MSF N tool recently. A major change has been changing from modelling the yield response inside the tool, to modelling yield inside the tool and then using smoothing splines to figure out the yield response to N rate.

The next version of the N tool will also incorporate likely spring temperature.

It should be ready for a 'beta' release in a month or so. 

These images are from version four of CGIAR-CSI's Shuttle Radar Topography Mission elevation data. Each image corresponds to a 'tile', and the biggest is about 6MB. The legend has been chosen for the Wimmera/Mallee, but extended at the top and bottom end to cover the other less important bits of south eastern Australia. Click on an image to download the corresponding tile:

For those seeking the broader view of gamma-radiometrics for the south east of Australia, these Google Earth images are for Minty (2009)'s  Radiometric map of Australia (First Edition):

The legends are, where possible, the same as for the Victorian data.

South Australian Mallee/Riverland users need to be wary of a large area of poor data quality along the Sunraysia Hwy to Nuriootpa, and along the Mallee Hwy to Peake. There is an outline of the area here. There seems to be some data, but it must have been taken at wider spacing (less precise).

A drawback (at present) of the data download page at www.malleefocus.com.au/data is that you can't easily see the 'big picture' of the data. I've had to develop some images of the Shuttle elevation and GeoSciences Victoria gamma-radiometrics data for some work on the BCG/Vic No Till Flexible Farming Project. These images may be useful for people trying to understand what variation at the farm level might mean. 

I've put the images into Google Earth '.kmz' files. These are the links to the images. Those on slow connections need to beware: some kmz files are up to 10Mb and will take a while to download:

  The images illustrate some of the features that make the gamma-radiometrics data useful, and also the east-west artifacts that users south of Swan Hill and east of Birchip need to be careful of. 

Soils around rivers and lakes have much in common with soils on hills to the south-east. The 'fingers' of soil blown in from South Australia's coast can be seen to be mixed in to some of the areas around them. The 'plains' of the southern Mallee and northern Wimmera are somewhere between.

The elevation data shows clearly the outline of the ancient Lake Bungunnia, below about 65m, and the possible past course of the Murray river along the now Wimmera river.  'Strand lines' (beach dunes) of the retreating Murrayvian gulf and lunettes of ancient lakes can also be distinguished from the dunes that are more obvious at ground level.

I'm not sure how to interpret the magnetics. It's supposed to relate to depth to bedrock. It does show that there's some weird things going on, centered on Culgoa and from Horsham northwest to Murrayville. 

If you would like to share an observation on the data and what you think it means, please post a comment! 

This is an example of what you will see in Google Earth, in this case the elevation image over the Tyrrell basin (it's upside-down - I was looking from the north):

I’ve been working (with Mallee CMA) on a web page that allows farmers to download publicly available spatial data: www.malleefocus.com.au/data

It’s been through various phases of testing and seems to work OK.

The aim of the page is to give Wimmera and Mallee farmers free access to some spatial data that will help them make variable rate/zoning decisions in time for sowing 2009, especially if they don’t have other sources such as yield maps or EM38 surveys. There are some good sources of spatial data out there – DSE have elevation on a 10 x 10m grid that covers the whole of the state, and GeoSciences Victoria have gamma-radiometrics on a 50 x 50m grid. They are publicly available but have been difficult for farmers to get and process in the past.

The web page allows farmers to draw their property on a Google map. After filling in a few details they will then receive an email containing ‘shapefiles’ of data for that area. The shapefiles can be opened in precision agriculture software such as SMS, Case AFS, John Deere’s Apex, and free GIS programs such as MapWindow. The elevation data will help to sort hills, flats and low spots. The gamma-radiometrics data relates to the parent material of the top 30cm of soil and may be useful for understanding soil variation in some areas.   

The work has been a ‘love job’ and farmers may have to be patient if the emails take a day or two to come through. I’d also be keen to have feedback on the site if anyone has ideas on how it could be improved. There is some data there for SA and NSW and I will add new data as I become aware of it. If you come across any, please email!

 This post contains some of the verification results (1997-2007)  from version 1.5.1 of the N tool. To appreciate the verifications, it's important to understand how they were derived.

The N tool contains a statistical 'model' of APSIM wheat yield response to nitrogen curves, so the verification is designed to 'verify' that the statistics produce output that is something like what was used to produce the model in the first place.

The summary data from each simulation (eg. Table 1 below) was fed into the N tool. Internally, the N tool used the EC1:5 and N rate directly, but 'grouped' the growing season rainfall, April 1 soil water, and sowing date, before extracting corresponding nitrogen responses from its internal model. This means, for example, that for the 2003 data in Table 1, GSR between 160 and 182mm are all treated the same (GSRGp = 3). April 1 sowing water between 30 and 41mm are treated the same (WatGp = 2). The sowing day is also broken into 10 day groups.

Table 1. An example of how APSIM data has been fed into the MSF N tool, for the 0 kg N/ha fertiliser rate, on Carwarp soils in 2003. The N tool takes the N rate directly, and the EC1:5 (average 0-110cm) as a summary of the soil type. The growing season rainfall (GSR) is put into 50mm groups, April 1 water 15mm groups, and the sowing day into 10 day groups.

Considering the simplified way the N tool handles input data, it seems to do a reasonable job of reproducing relative fertiliser responses between soil types and fertiliser rates (0-180 kg N/ha), within each year and weather file (Figures 1-11). Carwarp, Euston, Loxton and Pinnaroo denote different weather files for each year. Each point represents a weather file (site) - soil - fertiliser rate combination.

The difference between years and weather files is not reproduced as well (if it was they would all be on the same curve). This at least partly reflects grouping of the input data.

The years 2003 and 2000 stand out as having particularly poor relationships between the nitrogen responses the N tool estimates, and the original APSIM simulations, worse for some sites (eg. Pinnaroo, Euston) than others (eg. Carwarp, Loxton). Ultimately it may be necessary to introduce other factors (eg. 'dry spring') to improve the results from such a simple statistical model. 

The N tool also tends to over-estimate low nitrogen response rates, and under-estimate high response rates in some years (eg. 2005). This is also something that needs to be investigated to improve the N tool.

Figure 1. 2007 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing. 

Figure 2. 2006 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 3. 2005 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 4. 2004 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 5. 2003 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 6. 2002 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 7. 2001 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 8. 2000 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 9. 1999 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 10. 1998 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

Figure 11. 1997 verification, yield response to nitrogen estimated by the MSF N tool 1.5.1 compared with the APSIM simulated yield response on which the N tool is based. Each site is a different weather file, one point is a nitrogen response on one of 12 soils for one of 0, 7.5, 15, 30, 60, 90, 120 and 180 kg N/ha nitrogen applied at sowing.

This version of the N tool has an improved internal model of the nitrogen response.

I have made the sowing dates between 18 Jun and 8 Jul into one group, put years with greater than 300mm GSR into one group, and put April 1 water > 90mm into one group. This makes the statistics a bit more robust and fixes the Pinnaroo late sowing problem identified by Fiona Best. I also discovered that the model did a better job with a logarithmic model of EC1:5 (instead of linear), so that's incorporated too.

While doing (much) analysis to try to improve on what was there, I experimented with ways of representing poor rainfall distribution eg. dry spring, dry winter. This definitely has an impact on APSIM and would be good to include in future. 

Looking at validations of the model inside the N tool (against the original APSIM simulations), it doesn't do a bad job in some years but there are some year-location combinations where it is not too good. I'll post some of this in future.

I have just uploaded version 1.5.0 of the MSF Zonal N Management tool.

The new version includes some improvements to the optimisation options suggested by some users (John Stuchbery):

- the threshold where the tool stops applying nitrogen is now a $return:$spent ratio (Economic settings)

- the tool can be set to apply nitrogen until the threshold is reached, or until an average amount of nitrogen has been used

- the tool can be set to redistribute small nitrogen amounts where profit < cost of application

The user interface has been changed so zones and paddocks can be minimised, and deleted using the familiar 'X' button (David Gobbett). Some unnecessary repainting has been stopped to speed it up.

The help has been improved to give help about specific options in the tool (rather than whole components of the tool eg. 'paddocks').

The next issue to tackle will be improving the statistical model behind the nitrogen responses. We've now compared them to the original APSIM ones and they're not bad, but definitely room for improvement, and fixing some problems (eg. the Pinnaroo late sowing problem pointed out by Fiona Best).

I did a run-through of the tool yesterday with Fiona Best of BCG, and noted some weirdness.

Firstly, what I have called the 'sowing date' is really the date of the break - when APSIM records enough rainfall to sow the crop.

At Pinnaroo, sowing days in the 28 Jun - 8 Jul range have a very high N response (compared to earlier or later), even with low GSR (the 100-150mm range is most obvious). The reason for this is that most of the crops in the analysis with this sowing day come from 1986 - a year with 300-350mm GSR. The analysis depends on having a number of years in each sowing date range to give sensible numbers and 1 (and a bit of 1957 too) is not enough!

To solve it I might first try combining 18 Jun - 8 Jul into one big sowing date category. There are lots of years in the later 8 Jul - 18 Jul category because this is when APSIM decides to sow a crop regardless of the 'break'.

The tool deals with fertiliser at sowing only, hence the application cost should probably default to $0/ha (instead of $10/ha in the demo paddock). Application cost isn't considered in the optimisation process so it's possible to end up with a negative profit in zones that only get small N rates, even though the marginal response to nitrogen is profitable. It should be an option to redistribute small N amounts that don't pay back the cost of applying them.

I've also had some comments from John Stuchbery. The economic 'loss threshold' is in $ profit/kg N applied. It needs to be adjustable in smaller units than '1'. If you are looking for a 1:1 return on average then it should be set to $0/kg N. For a 2:1 return, it should be set to the N price, eg if N is $1.50/kg then the 'loss threshold' should be $1.50/kg profit. Alternatively it might make more sense to make the 'loss threshold' a ratio (eg 2:1 return) instead.