1. PROJECT TITLE
QPI20 Development of a National Drought Alert Strategic Information System (May-1996)
2. VOLUME 6: WHEAT MODELLING SUB-PROJECT
3. CONTACT DETAILS:
a) Primary Research Organisation: Climate Impacts and Applications
Resource Sciences Centre,
Queensland Department of Natural Resources
80 Meiers Road, Indooroopilly, 4068.
b) Principal Investigator: Mr Ken D. Brook
Principal Scientist
Tel: 07-3896 9379 Fax: 07-3896 9606
4. DPI research staff:
John Carter, Tim Danaher, Greg McKeon, Cheryl Kuhnell, Neil Flood, Graeme Hammer, David Butler, Robert Hassett, Helen Wood, Alan Beswick, Alan Peacock, Colin Paull, Patricia Hugman
5. Collaborating Organisations
(i) Research Organisations
Greg Beeston, Greg Mlodawski, David Stephens, Agriculture Western Australia.
Dennis Barber (now NSW DLWC), Russell Flavel, Department of Environment and Land Management, South Australia
Rik Dance, Danny Brock, Don Petty, Department of Primary Industry and Fisheries, Northern Territory
Daryl Green, David Hart, Rob Richards, Department of Land and Water Conservation, New South Wales
(ii) Funding Organisations
Land and Water Resources Research and Development Corporation
Grains Research & Development Corporation
Goodman Fielder Mills Ltd.
6. FOR FURTHER INFORMATION:
There are 6 volumes of documentation available on LWRRDC QPI20. The volumes are:
(1) Research summary
(2) Field validation of pasture biomass and tree cover
(3) Development of data rasters for model inputs
(4) Model framework, Parameter derivation, Model calibration, Model validation, Model outputs, Web technology
(5) Evaluation of model performance relative to rainfall, Inter-state model calibration, Extension, and State comments.
(6) Wheat modelling sub-project. (this document)
A short video of computer visualisations produced from the project is also available.
The above information sets are available at a nominal charge to cover printing and distribution costs.
For more detailed information contact:
Spatial rangeland model, drought alerts Mr John Carter, Resource Sciences Centre, Queensland Department of Natural Resources, 80 Meiers Rd, Indooroopilly. 4068.
Ph 07 - 3896 9588 Fax: 07 - 3896 9606
GRASP pasture simulation Dr Greg McKeon, Resource Sciences Centre, Queensland Department of Natural Resources, 80 Meiers Rd, Indooroopilly. 4068.
Ph 07 - 3896 9548 Fax: 07 - 3896 9606
Meteorological data Mr Neil Flood, Resource Sciences Centre, Queensland Department of Natural Resources, 80 Meiers Rd, Indooroopilly. 4068.
Ph 07 - 3896 9734 Fax: 07 - 3896 9606
Wheat simulation modelling Dr Graeme Hammer, Agricultural Production and Systems Research Unit, Queensland Department of Primary Industries, Tor Street, Toowoomba. 4068. Ph 076 - 314 379 Fax: 076 - 332 678
Wheat statistical modelling and yield forecasting Mr David Stephens, c/- Agriculture Western Australia, Ngala Annex, Baron - Hay Court, South Perth. 6151.
Ph 09 - 368 3983 Fax: 09 - 368 3946
Evaluation of model performance relative to rainfall, Inter-state model calibration, Extension, and State comments / Page iiiLWRRDC QPI 20 / Volume 6
SECTION (a) - Development of Predictive Models of Wheat Production
Contents
1. PROJECT TITLE
2. VOLUME 6: WHEAT MODELLING SUB-PROJECT
3. CONTACT DETAILS:
4. DPI research staff:
5. Collaborating Organisations
6. FOR FURTHER INFORMATION:
7. ABSTRACT
8. INTRODUCTION
9. MATERIALS AND METHODS
9.1 Study Period and Data Sets
9.2 Rainfall Regression Model
9.2.1 Location
9.2.2 Time
9.2.3 Monthly Rain
9.3 Weighted Rainfall Index
9.4 Agro-climatic Models
9.5 Stress Index Model
9.6 Drought Index
9.7 Calibrating the Indices - Trends in Wheat Yields
9.8 Common Water Balance
9.9 Simulation Models
9.10 TACT Simulation Model
9.11 APSIM-Wheat
10. RESULTS AND DISCUSSION
10.1.1 Individual Shire Results
10.1.2 State / National Averages
10.2 Implications
11. CONCLUSIONS
12. REFERENCES
List of Figures
Figure 1: 127 rainfall stations used by the agro-climatic models; + = 70 stations with temperature data
Figure 2: Australian cropping boundaries and rainfall stations used in the multiple regression model.
Figure 3: ABS yield (t/ha) for 1982
Figure 4: ABS yield (t/ha) for 1983
Figure 5: Stress Index yield (t/ha) 1982
Figure 6: Stress Index yield (t/ha) 1983
Figure 7 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) APSIM-Wheat simulation for the Waggamba shire.
Figure 8 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) APSIM-Wheat simulation for the Murilla shire.
Figure 9 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) APSIM-Wheat simulation for the Banana shire.
Figure 10 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) APSIM-Wheat simulation for the Bungil shire.
Figure 11 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) TACT simulation for the Merredin shire.
Figure 12 Comparison between observed yield (ABS) and predicted yield for (1) rainfall regression, (2) stress index, (3) drought index, and (4) TACT simulation for the Cunderdin shire.
Figure 13 Comparison between observed (ABS) and predicted (weighted) Queensland yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index.
Figure 14 Comparison between observed (ABS) and predicted (weighted) Queensland yields for the APSIM-Wheat model
Figure 15 Comparison between observed (ABS) and predicted (weighted) New South Wales yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index.
Figure 16 Comparison between observed (ABS) and predicted (weighted) Victorian yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index.
Figure 17 Comparison between observed (ABS) and predicted (weighted) South Australian yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index
Figure 18 Comparison between observed (ABS) and predicted (weighted) West Australian yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index
Figure 19 Comparison between observed (ABS) and predicted (weighted) Australian yields for (1) rainfall regression, (2) stress index, (3) drought index, and (4) weighted rain index
List of Tables
Table 1 Summary of models used to estimate yields
Table 2 Maximum monthly rainfall amounts allowed in model calculations for different regions and periods.
Table 3 Yield trends (t/ha/year) for (a) a period where yield growth had levelled off, and (b) recent intervals chosen to represent more recent trends (excluding sequences of extreme years), with ** P<0.01
Table 4 Criterion to calculate the midpoint of district sowing.
Table 5 R2 and MAE (t/ha) values from a fit of observed (ABS) vs predicted values for each method at 6 selected shires.
Table 6 R2 and MAE (t/ha) values from a fit of observed weighted yield (ABS) vs predicted weighted yield for each method for each state and nationally.
Wheat Modelling Sub-project / Page iiiLWRRDC QPI 20 / Volume 6
7. ABSTRACT
In developing a wheat yield forecasting capability for a National Drought Alert Strategic Information System, a feasibility study of six yield forecasting models (rainfall regression, weighted rainfall index, stress index, drought index, TACT simulation model and the APSIM-Wheat simulation model) from three classes (empirical, agroclimatic and simulation) was undertaken and evaluated at a number of scales. Essentially it was found that while all classes have the potential to satisfactorily forecast Australian wheat yields at the shire scale, the simpler empirical and agroclimatic models performed best overall, particularly when the demand on computing resources is taken into account. Of these, the empirical approaches were marginally superior to the agroclimatic models in a number of cases and equivalent in others. It should be noted, however, that poor knowledge of some input data layers did impact on the performance of the crop simulation models; efforts to overcome this would need to be balanced against the expected gain in precision given the performance of the simpler models.
8. INTRODUCTION
Australia is a major exporter of wheat and coarse grains. Wheat production alone is worth in excess of $2000 million per annum (AWB, 1993), but this varies as a result of one of the most variable climates in the world (Russell, 1988). Large annual fluctuations in yields (and production) are of major concern to marketing agencies who have to sell this grain on a volatile world market. While other grain exporting countries have developed systematic techniques and models to forecast crop yields (Motha and Heddinghaus, 1986; Stephens, 1988; Walker, 1989), Australia has yet to use any formal forecasting procedure to estimate its own production (AACM, 1991). Existing forecasts are based on a compilation from various sources of information and are unreliable and only broadly indicative (AACM, 1991).
To address this problem, this project sought to:
1. develop and apply approaches capable of forecasting Australian wheat yield and its spatial distribution at the shire scale by considering approaches covering a range in complexity, and
2. compare the predictive ability of the approaches and assess the trade-offs between accuracy and likely cost of application in a real-time forecasting mode.
A vast number of crop weather models have been proposed and these have been reviewed in the Australian context by a number of authors (McMahon, 1983; Rimmington et al., 1986; Angus, 1991; AACM, 1991). Models that can specifically forecast yields have been classified into three broad categories (Baier, 1977; 1979):
1. Empirical models that relate weather (typically rainfall) and soil variables directly to yield through a statistical model.
2. Crop weather analysis models that evaluate crop response to variations in derived agrometeorolgical indices that are usually based on a simple water balance; and
3. Simulation models - that explicitly model plant growth and development with a set of mathematical equations which have a physical, chemical and physiological basis.
The development of simulation models of ever increasing complexity has been a feature of research in the last two decades, but there are now signs of real efforts to return to simpler and more general models (Nix, 1985). The increased cost, effort and complexity of more recent crop models have not lead to commensurate improvements in predictions (Norman, 1981). Complex models have often been significantly less successful at simulating grain yield (on a broad scale) than more empirical models that have been tested locally (White et al., 1993). However, due to the short history of the field and the difficulty in obtaining adequate data sets, no direct comparisons of model accuracy has been made for different levels of model complexity.
At the simplest level, Lehane and Staple (1965) found in Canada that multiple regression equations based on rainfall distribution and soil moisture gave better estimates of yield than did equations based on total seasonal rainfall alone. Similarly, Nix and Fitzpatrick (1969) showed an agrometeorological (stress) index based on soil water supply at ear-emergence considerable improved on equations based on total crop rainfall, soil moisture at sowing and sums of evapotranspiration. Hashemi (1976) however found that a simple moisture balance did not improve on growing season rainfall totals in Iran.
Unfortunately, there is a dearth of information on model comparisons as modelling progresses from water based budgets to simulation models of plant growth. This report therefore reviews the utility of various yield forecasting models deemed appropriate for Australian conditions. Modelling approaches are listed in Table 1.
Table 1 Summary of models used to estimate yields
Class / Model / DescriptionEmpirical / Rainfall Regression / Empirically derived relationship between historical shire yields and rainfall
Weighted Rainfall Index / Linear function of (biologically based) weighted monthly rainfall for Australian rainfall districts. (Stephens et al., 1994) The derived index is calibrated against historical yield data.
Agroclimatic / Stress Index / An index is derived from water stress relative to plant available water using daily rain, average weekly temperature and radiation data throughout the growing season. (Stephens et al., 1989) The index is calibrated against historical yield records.
Drought Index / Calculates a seasonally integrated daily growth index from a physiologically based model using daily rain and temperature. (Walker, 1989) As with the stress index the drought index is calibrated against historical yields.
Simulation / TACT / A derivative of the CERES-Wheat model. (Robinson and Abrecht, 1994)
APSIM-Wheat / Woodruff-Hammer simulation model (Hammer et al., 1987)
Of these models, most can be run at a shire (or lower) level and predictions scaled up by weighting to a forecast at a national scale. The exception is the weighted rainfall model which runs on district rainfall and can only be aggregated to forecast state or national yields. Apart from the Weighted Rainfall Index, the basic unit for predictive purposes was the shire (or county in South Australia). No claim is made that any one model is the definitive in its class but rather they are representative of their type. An exhaustive search, particularly of the empirical class, was beyond the scope of the project.
Due to data and time constraints this study could only test the simulation models at two shires in Western Australia (TACT model) and four shires in Queensland (APSIM-Wheat model). To aid in model comparison the same fallow water balance (Ritchie, 1972) was utilised (from the CERES-Wheat model - Ritchie and Otter, 1985) in both the Crop weather analysis models and the Tact simulation model. A modified form of this is also used in the APSIM-Wheat model.
9. MATERIALS AND METHODS
9.1 Study Period and Data Sets
In the process of model development many years of data are needed to cover the full range of conditions and derive the necessary parameters. King (1989) suggests at least twenty are needed for multiple regression models, if too few are used relative to the number of predictors then overfitting can be a problem.. The Agroclimatic index and simulation models have a less strict requirement as only one parameter or index is being regressed against historical yields. However, in all cases an extended range of years is essential to avoid spurious results from runs of advantageous seasons.
In this study, the most recent 18 years of available yield data was chosen as this has a good mixture of drought, average and wet years and yield fluctuations were the greatest on record (Hamblin and Kyneur, 1993). This period was also chosen because older crop varieties were replaced from the mid 1970's with varieties that have benefited from the successful incorporation of the reduced height (dwarfing) and multiple rust resistance genes (Zwer et al., 1992; Hamblin and Kyneur, 1993). Present varieties have a superior resistance to lodging, higher yield potential, higher harvest index and less incidence of rust (Richards, 1992; Zwer et al., 1992). Higher yields closer to potential are now possible in wetter years when yield losses are highest.