Can imperfect seasonal climate forecasts improve farm profit

Background: developments on seasonal climate forecasts

More than 120 years ago, Sir Charles Todd (in 1893) stated that “the importance to the farmer, the horticulturalist, and pastoralist of knowing beforehand the probabilities of dry or wet seasons, and whether the rains will be early or late, or both, has naturally led to a desire for seasonal forecasts. They have them, it is said, in India; why not in Australia?”

About 20 years ago, two of the pioneers of the use of seasonal crop forecasts (SCF) in Australia, Hammer and Nicholls commented that “we are confronted with unprecedented opportunities to tune our agricultural systems - we have a seasonal forecasting capability. We have started to think through how we can best use the knowledge that the next season is not a total unknown.”

Uncertainty about the coming season makes planning and decision in rain-fed agriculture difficult. It follows that any information on the coming season that is better than guessing, is potentially valuable. However, communicating and using seasonal climate forecasts has been more difficult than first thought. 

It is useful for farmers and advisers to understand the basis of seasonal climate forecasts. It is essential to think through how to use seasonal climate forecasts in decision making.

Statistical seasonal climate forecasts are based on historical relationships between rainfall at a location and sea surface temperature patterns (e.g: El Nino) or the Southern Oscillation Index. Dynamic seasonal climate forecasts require powerful computers to model future states of the atmosphere and ocean. In Australia, the Bureau of Meteorology has shifted from a statistical model to a dynamic model called POAMA.

The next dynamic seasonal forecasting model will be ACCESS-S based on ACCESS (Australian Community Climate and Earth System Simulator), already in use for seven-day weather forecasts and climate change projections. The last “S” in ACCESS-S stands for seasonal. The Bureau of Meteorology and CSIRO expect ACCESS-S to show significant improvement over POAMA. POAMA works on a grid of about 250km, whereas ACCESS-S will use 60km. 

The remaining portion of this paper reports on a study using POAMA for top dressing at Hart in South Australia. This is based on Hayman et al 2015 at the Australian Agronomy Conference held in Hobart.

Valuing seasonal forecasts: a case study at Hart, South Australia

The use of N fertiliser is a major source of yield and profit for Australian grain farmers. At the same time, N fertiliser represents a significant cost and hence a source of risk when crop demand is less than expected. N fertiliser can be a pollutant through losses as runoff and leaching. Recently there has been increased attention to nitrous oxide as a potent greenhouse gas. The need to maintain and increase profitability of N while reducing the economic and environmental risk requires a knowledge-intensive approach that takes into account the supply of N from the soil and fertiliser and the demand of N from the crop. 

Growing season rainfall (GSR) explains most of the year to year variation in crop demand. Seasonal climate forecasts have been available in Australia since the early 1990s. There have been many studies on the use of SCF on N decision making in wheat. Early studies such as Marshall et al 1996 and Hammer et al 1996 focussed on the grains industry in Queensland using phases of the Southern Oscillation Index. In a recent study Asseng et al 2012 reported on the dynamic climate model POAMA for N on wheat in West Australia. 

This study set out to investigate the value of POAMA for a simple simulation exercise of top dressing decision on wheat in South Australia. As with any analysis, the conclusions are sensitive to the assumptions.

Methodology: how we set up APSIM and analysed the output 

For a wheat crop with a given level of starting soil N, the response to top-dressed N in early August (a common farm practice) is determined by both the climate up to that point and the climate for the rest of the season. In this simulation exercise we focussed on the season finish and sought to avoid confounding the results with a range of starts to the season. To achieve this, 30 daily climate files were created using the same data from January 1 to August 10, but with different data for each year for August 11 to December 31. This raises the question of what daily data to use for January 1 to August 10. A single year is preferable to any averaging of daily data, but it is not obvious which year to select. Simulation runs with APSIM showed that a combination of biomass and soil water on August 10 were good indicators of final yield and N response. Probability distributions of biomass and soil water on August 10 were generated using a mid-maturity wheat cultivar sown each year on May 15 into a soil characterised in APSIM (light clay over medium clay over heavy clay (Hart No284)) with a Plant Available Water Capacity of 183mm. The initial soil water, soil N and surface organic matter were reset on January 1 every year. The soil N was set to 60kg/ha in the top 60cm, plus a further 12kg/ha at lower depths. The surface organic matter was set at 1000kg/ha with a C:N ratio of 80. 

From the probability distributions of biomass and soil water we selected a year that represented below average (1.1t/ha biomass and 45mm extractable soil water (ESW)), average (1.6t/ha biomass and 73mm ESW) and above average (2t/ha biomass and 115mm ESW). In this paper we are reporting the results for the below average start. For each of the 30 years, we simulated yield, leaching and de-nitrification as a result of top dressing 0 to 300kg of N. The South Australian gross margin handbook was used to provide the five-year average of ASW wheat at $270/tonne, a growing cost (excluding N) of $190/ha, and a nitrogen fertiliser cost of $1.30 per kg.

We accessed gridded rainfall forecasts from POAMA-2(M24) for the 30-year period 1981 to 2010 and used simple spatial interpolation to obtain station level data for Brinkworth. An above/below median forecast was recorded for a year if the mean of the 33 ensembles for that year was above/below the median of the ensemble means in the other 30 years. When compared with historical data for Brinkworth the forecast was correct (eg: below median when below median was forecast) 19 years and incorrect 11 years.

Results and discussion

As shown in the All Years response in Figure 1(a), top dressing with N lifts simulated yields from 1.5t/ha to 2.7t/ha (about 80 per cent). Separating the response into above and below median rainfall years highlights the strong influence of August to October rainfall. This is evident in both the agronomic response curves (Figure 1a) and consequent economic response curves (Figure 1 b). The comparison of agronomic and economic responses to N in Figure 1a and b shows that when the cost of N is taken into account, the highest gross margin will be at a lower rate than the highest yield. 

While it is obvious that adding N past the maximum yield wastes money, using the average response curve (shown in Figure 1b) leads to over-fertilising in below median seasons and under-fertilising in above median seasons.

Two line graphs showing figure 1, average simulated response of yield and gross margin to fertiliser rate

Figure 1: (a) average simulated response of yield and (b) gross margin (right panel) to fertiliser rate for 1981 to 2010 (All Years), average of the 15 below median years (BM) and average of the 15 above median years (AM). The above and below median years are determined by rainfall from August 11 to the end of October. The assumptions for the gross margin are detailed in the methods.

Figure 2 reinforces the highly variable year-to-year response to top dressed N. In a high-yielding year like 1992, shifting from 10kg/ha to 70kg/ha almost doubled the gross margin from $405/ha to $725/ha. 

This is due to a simulated yield increase from 2.4 t/ha to over 4t/ha.   In other years, such as the El Nino drought of 2006, the simulated yield was less than 1t/ha and adding N only increased the loss.

A line graph showing figure 2, time series of simulated gross margin and marginal return for topdressing.

Figure 2: (a) time series of simulated gross margin and (b) marginal return (right) for topdressing with 10, 40 and 70 kg/ha. The marginal return graph for 40 kg/ha is the extra profit ($/ha) gained by increasing the rate from 30kg/ha to 40kg/ha divided by the cost ($/ha) of adding the extra10kg/ha. 

Figure 2b shows that the return to investing in N top dressing fluctuates from year to year. As expected, the first 10kg/ha of N will always have a much higher return than any further increase, but there are years when the return for even the first 10kg of N is less than one dollar for a dollar invested.

Farmers should hope to at least receive $1.10 for each $1 invested. Higher rates greatly increase the risk of the return falling below $1.00 or $1.10.

Figure 3 shows the declining marginal return with increased rates of N. Tracking the “All Years” line in Figure 3 shows that on average, the marginal return over the 30-year period for the first 10kg N/ha is about $4. The return for shifting the top dressing rate from 30kg to 40kg N/ha is $1.57 and for shifting from 60kg to 70kg N/ha is $0.89. Figure 3 also shows the impact of spring rainfall. The above-median curve drops below the $1.10 return in the step from 90kg to 100kg/ha. In contrast, in the below-median years, even the step from 20kg to 30kg/ha drops below this line.

A line graph showing figure 3, the declining marginal return with increased rates of N.

Figure 3: Marginal return for each kg of N assuming that wheat is $2.50/kg and N is $1.30/kg.

The points on the graph (Figure 3) are simulations for each 10kg of N top dressed per hectare from 10 to 175, then 25kg/ha steps from 175 to 200 and 50kg/ha steps from 200 to 300. As per Figure 1, All Years refers to the 30 years from 1981 to 2010 and above and below median (AM/BM) refer to the 15 years in the sample when the rainfall from August 11 to the end of October was above or below median.

A portion of the N lost through de-nitrification is likely to be lost as nitrous oxide. Figure 4a shows that this loss is primarily driven by the rate of fertiliser. At higher rates (150kg N/ha to 300kg N/ha there is a strong interaction between rate and season. The large de-nitrification losses come from a small number of years as shown by the time series (Figure 4b). 

Two line graphs showing figure 4, de-nitrification losses and time series of de-nitrification losses.

Figure 4: (a) de-nitrification losses as a response to top-dressed fertiliser averaged over 30 year period 1981 to 2010 (All Years) and the 15 years with above and below median rainfall for 11 August to end of October. And (b) time series of de-nitrification losses for 10, 40 and 70 kg N/ha. 

Figure 5 compares the outcome over the 30 years in term of risk as measured by the percent of times that the marginal return was less than $1.10 and return as measured by the gross margin. The lower arc of points represent fixed rates of N from zero up to 170kg N/ha. As shown in Figure 1(b), the gross margin will increase up to 100kg of N per hectare. Higher rates decrease the gross margin. The line running from zero to 100kg N/ha is an efficiency frontier as the attainable points are below it, and each point on the frontier is to the left (lower risk) or above (higher return) of any other attainable point. The higher arc of points represent decision makers who are following the forecast and adjusting their N according to the rule shown in the figure.

A line graph showing figure 5, risk and return of topdressing with and without a forecast.

Figure 5: Risk and return of topdressing with and without a forecast. 

The lower arc of points in Figure 5 are the 30-year average return from using N rates and the upper arc of points are the average risk and return from making an adjustment based on the forecast. The lines joining the points represents the efficiency frontier without the forecast (lower line) and with the forecast (upper line). The movement between points a, b and c are explained in the text below.

The main message of Figure 5 is that POAMA forecasts can shift the efficiency frontier. A farmer can change from applying 30kg/ha every year (point a) to point (b) applying 10kg when below median is forecast and 50kg when above median is forecast. This leads to a gain of $8/ha and a reduction in risk (as measured by percentage chance of <$1.1 return) from 35 per cent to 31 per cent. 
Points (a) and (b) represent the same amount of fertiliser on average (30kg/ha). 

In Figure 5, 40kg/ha is applied as 20 in below median and 60 in above median, and 50kg/ha is split 20:80. We calculated all combinations (for example 50kg/ha - 10:90, 20:80, 30:70 and 40: 60) and then selected the most efficient (highest return and lowest risk). 

It is interesting that at higher rates, the optimum level for below median is 50 or 60kg N/ha. This could seem to contradict Figure 3 that shows for below median rainfall farmers should not apply more than 20kg/ha, because the marginal return for the next step to 30kg/ha drops below $1.10/ha. However, this logic would only hold if the forecast from POAMA was a perfect forecast. The shift from point (a) to point (c) represents the use of extra fertiliser with the forecast. This results in an increase in gross margin of $23/ha. It is important to note that some of the benefit of the “with forecast” case comes from the information being used to apply extra fertiliser.

Another version of Figure 5 could be drawn with the x axis as environmental risk as measured by the nitrous oxide emissions or N leaching. Under the assumptions of this simulation study, there was little change to nitrous oxide emissions or N leaching from a farmer following or ignoring POAMA forecasts. The main reason for this is that the economic optimum supply of N is below the maximum crop demand in average to good conditions.  

Conclusion

Considering the marginal economic response will focus attention on the most efficient use of N. This will have the side benefit of reducing the chance of excess N increasing environmental risk.

References

Asseng S, P McIntosh, G Wang and N Khimashia (2012). Optimal N fertiliser management based on a seasonal forecast. European. Journal of Agronomy, 38, 66-73.
Hayman, P., B. Cooper, K. A. Parton, O. Alves, G. Young, H. BH and C. Scheer (2015). Can advances in climate forecasts improve the productive and environmental outcomes from nitrogen fertiliser on wheat? A case study using POAMA for topdressing wheat in South Australia. 17th Australian Agronomy Conference, Hobart. http://www.agronomy2015.com.au/proceedings
Hammer, G. L. and N. Nicholls (1996). Proceedings of the 2nd Australian Conference on Agricultural Meteorology. Brisbane, Australian Bureau of Meteorology: 19-27
Hammer G, Holzworth D and Stone R (1996) Aust. J Agric Research 47,717-37
Marshall G, Parton K and Hammer G (1996) Aust J Agric Economics 40, 211-33

Acknowledgements

This project was funded through the Commonwealth Government Filling the Research Gap component of the Carbon Farming project titled “Can advances in mid-term weather forecasts reduce emissions from nitrogen fertiliser?” 

The Managing Climate Variability Program was a co-investor in this research. The GRDC is the major contributor to the Managing Climate Variability Program. 

Contact details 

Peter Hayman 
SARDI Climate Applications
GPO Box 397, Adelaide, SA
0401 996 448