Predicting regional grain sorghum production in Australia using spatial data and crop simulation modelling

Abstract

Grain sorghum is the major dryland summer crop produced in the subtropical region of Australia. Production variability is great and the consequent uncertainty about likely production restricts marketing options and contributes to instability in grain price. Improved methods for predicting regional production would assist marketing decisions for farmers and grain traders. The objective of this study was to determine whether reliable regional grain sorghum production predictions could be generated by combining crop simulation and geographic information system technologies. We used historical shire production data to test the approach using hindcasting. Geographical data bases of landscape and soil attributes were used to define arable land boundaries and soil properties. Geographical data bases of daily rainfall and climate were overlaid and used to drive a spatial simulation of sorghum production for all shires in Queensland for the period 1977–1988. The results of the simulation were compared with production statistics at the shire and aggregate state levels. The spatial integrity of the prediction system was examined by comparing maps of predicted and reported shire production for specific years. There was a general tendency for the simulated yields per unit area to be greater and more variable (from year to year) than the historical shire production data. This probably reflected the fact that the simulation assumed perfect management and pest-free conditions. The relatively coarse spatial interpolation of rainfall would likely also contribute to this outcome. Linear regression relationships were developed between historical and simulated data at the shire scale to calibrate the simulated yields. Estimates of total production for each shire (in any year) were derived from the predicted yield per unit area, which was derived from the regression correction of the simulated yield, and the reported area planted. Excellent agreement between predicted and reported production occurred both at the individual shire and aggregate state (r = 0.96) scales. This approach was compared with use of mean shire yield as the estimate of predicted yield per unit area to examine the contribution of the yield simulation procedure to production prediction. Significant improvement in production prediction was attributed to the yield simulation. The spatial distribution of shire production estimates was examined by categorising and mapping shire production predictions. Comparisons with reported production estimates showed the integrity of the spatial distribution was largely retained. Hence, we conclude that reliable shire and state sorghum production estimates can be generated by combining crop simulation and geographic information system technologies. The procedure is suitable for further development for use in real-time. By updating estimates as a season progresses, improved timeliness and accuracy of production forecasts could be achieved.