Development of combined seasonal weather and crop productivity forecasting systems

Andrew Challinor¹, Julia Slingo¹, Tim Wheeler² and David Grimes¹
Departments of: ¹Meteorology ²Agriculture, University of Reading. U.K.

1. Introduction

Background

In order to predict crop productivity on seasonal time-scales, an understanding of both crop development and seasonal weather forecasting is required.

Fig. 1. A groundnut in flower.

Simple relationships between meteorological predictor variables and crop yield are pragmatic but not necessarily robust (see figure 2). This point is particularly relevant when considering possible future climate scenarios.

Fig. 2. Groundnut yield and precipitation, R²=0.52, although this correlation visably deteriorates in the latter half of the plot.

Employing an inter-disciplinary modelling approach to design an integrated predictive system should increase both understanding and accuracy. Challenges include:

The discrepancy in spatial scale between General Circulation Models (GCMs) and crop models.
Uncertainty in crop model inputs such as soil type, genetic coefficients, management practices etc.
Errors in GCM output.

Aims

Specific aims of the project are:

Assessment of the skill of seasonal forecasts in providing the parameters necessary for crop simulations.
Identification of potential shortcomings in the meteorological and crop numerical models.
Development of a probabilistic approach to crop productivity forecasting
Design of an integrated system for the prediction of seasonal weather and crop productivity for the seasonally-arid tropics.

The initial focus of the project is on India, where good historical rainfall records exist, together with groundnut yield time series recorded on a district level (see figure 3).

Fig. 3a. Groundnut yield for India in 1993. Data from the International Crop Research Institute for the Semi-Arid Tropics.

Fig. 3b. Mean groundnut growing area the the period 1996-1995. (ICRISAT)

Fig. 3c. Mean groundnut yield for India 1996-1995 (ICRISAT).

Factors determining crop yield

Fig. 4. Normalised standard deviation of a) groundnut yield and b) precipitation¹ for the period 1966 to 1995. Mean values are used for normalisation.

Fig. 5. Groundnut yield and precipitation for two Indian sub-divisions. West Gujarat is on the western-most extreme of India and West Madhya Pradesh is the second region east of this.

Management practices such as irrigation (see fig. 6) and choice of genotype are another important factor. It is important to be aware of changing practices when interpreting crop yield data and weather impacts.

Fig. 6. The irrigated fraction of the all-India groundnut yield.

3. Prediction of crop development

The CROPGRO model² is a well-established crop development model. Figure 7 shows the meteorological inputs required to run the model. In addition to these daily data, windspeed and humidity may be supplied, enabling the calculation of evaporation. Further inputs required include: plot spacing, genetic coefficients, (determined by calibration), soil type (hydrological properties, depth of soil layers etc.) and fertiliser application details. Sample model output is also shown in figure 7, with the coloured lines showing the effect of reducing rainfall by one climatological standard deviation for the whole period (blue line) and for a 25-day period (days 25-49) from flowering onwards (green line).

Fig. 7. The CROPGRO model inputs (upper panel) and output (lower panel) for the 1994 summer growing season in Hyderabad, India (17^oN, 79^oE).

Robust prediction of crop development requires a model detailed enough to account for important processes such as high temperature stress (figure 8), yet simple enough for the input requirements to be reasonable.

Fig. 8. CROPGRO model (open circles) and controlled experiment results (filled circles) for a six-day +10^oC episode at the time of flowering³

4. Combining weather and crop prediction models

Approach -

The first task is to identify the principle factors determining crop yield for the combined model. A balance must be struck so that predictions consider an optimum spatial scale - small enough to account for regionality, large enough for the results to be useful.

Once the appropriate level of complexity for the crop model has been determined, it can be tested with measured weather data if available, and/or model re-analysis data.

The gap in spatial scales between the two models must then be bridged. Crop models may be up-scaled via the use of meta-models, and GCM output can be down-scaled using statistical regression, nested models, or weather classification and re-sampling.⁴

Figure 9 summarises some of the scale issues. Not only do crop models and General Circulation Models operate at very different scales, but the plot yield data used to calibrate crop models is essentially point-data, whereas yields are often measured on the district level, so that many highly regional effects are averaged out.

Fig. 9. Schematic diagram of scale issues.

The challenge for a successful yield forecasting system is to combine meteorological and agricultural expertise at the interface of the two models.

More information and current work can be found here

References

1 G. B. Pant and K. Rupa Kumar, 1997. Climates of South Asia. John Wiley and Sons, Chichester, 320pp.

2 Boote, K.J., J.W. Jones, G. Hoogenboom and N. B. Pickering, 1998. The Cropgro model for grain legumes. pp99-128 in Tsuji et al (eds.), Understanding Options for Agricultural Production. Kluwer Academic Publishers.

3 Wheeler, T.R., P.Q. Craufurd, R.H. Ellis, J.R. Porter and P.V. Vara Prasad, 2000 (in press). Temperature variability and the yield of annual crops. Submitted to Agriculture, Ecosystems and Environment.

4 see e.g. Wilby, R.L and T.M.L. Wigley, Downscaling general circulation model output: a review of methods and limitations. pp530-548 in Progress in Physical Geography 21,4 (1997).