Crop model performance with seasonal prediction data

Posted on: 
Wednesday, May 13, 2015

Climate variability and extremes can have significant impacts on crops, so the ability to translate seasonal forecasts of meteorological variables such as temperature and rainfall into crop yield forecasts on a seasonal timescale has significant economic and humanitarian benefits. With this in mind, key tasks of EUPORIAS work package 23 have included the development of impacts models, their application to seasonal forecasts, and assessment of skill with respect to their accuracy for key sector specific variables such as crop yields. This report briefly summarizes general progress made in the work package relevant to the agricultural sector, and next describes the development, application and assessment of crop models for use with seasonal forecasts.

For most of the crop models used here (and most regions relevant to EUPORIAS), precipitation is the dominant driver of inter-annual variability in modeled crop yields. This implies that the models may be more confidently applied with seasonal forecasts in regions where precipitation predictability is strong. The importance of precipitation in driving modelled yields suggests that lengthy spin-up periods are not required, and that in some cases models can be reliably initialized with short spin-up periods, climatological forcing data on the sowing date, confirming previous work package findings.

A number of approximations may be made when driving the JULES-crop model for maize with WFDEI data without significant impacts on model performance in many regions, which are relevant to use with seasonal hindcasts, and likely also relevant to other crop models: a) Daily weather data may be used disaggregated to sub-daily values, if necessary; b) climatological weather data may be used, apart from for precipitation (which requires actualvalues); c), the model may be reliably initialized using WFDEI climatology on the sowing date.

Bias-correction of seasonal hindcasts to WFDEI climatology is important for crop modelling applications: as crop varieties are optimized to relatively narrow Thermal Time requirements, relatively small temperature biases may result in significant yield biases. The performance of the different crop models applied here varies with region, crop, forcing data and lead-time. For example, JULES_crop and GLAM perform generally poorly over East Africa for maize when driven by observed climate data, possibly because weather may not be the dominant driver of yield variability, due to inadequate parameter sets to represent local crop types, or due to missing processes (e.g. heat/water stress). Hence further model development and assessment is needed for confident application to seasonal hindcasts. Overall, model performance was generally best for maize in Kenya, and poorest for Ethiopia.  By comparison to simulations driven with reference forcing (WFDEI), LPJml showed generally good performance for maize when driven by seasonal hindcasts for the first rain crop, showing good ability to distinguish high, and particularly low productivity events, and better skill for shorter (1 month) lead times relative to longer lead times (3 months).


Availability of high quality, detailed observational datasets of crop yield and development is a significant challenge for both assessing and developing crop models for application to seasonal forecasts, particularly in East Africa. Given the potential value of seasonal crop productivity outlooks to the region, further effort should be invested in obtaining suitable datasets.