The use of probabilistic weather forecasts to predict crop failure

Process-based integrated modelling of weather and crop yield over large areas is becoming an important research topic. Our group in Reading has been developing a new approach to combined crop and climate forecasting. This modelling system couples crop simulation and numerical weather models on a common spatial scale based on observed weather / yield relationships. The development of a new crop model, the General Large Area Model for annual crops (GLAM), is a key part of this work. This has enabled simulation of crop yield on a regional scale using output directly from climate models and gridded observed weather data. These (deterministic) simulations have been shown to skillfully reproduce crop yields over large areas, in particular in western India, where there is a strong climate signal in observed yields.

In this presentation, the use of GLAM within a probabilistic framework is explored. General Circulation Models (GCMs) can be used to predict weather and climate months in advance by creating an ensemble of simulations. Each simulation has slightly different initial conditions and over a sufficiently long time period (two weeks and above) the results of each simulation can differ greatly. Since all of the initial conditions used are plausible, and within observational uncertainty, the ensembles contain probabilistic information.

Hindcasts of crop (groundnut) yield were produced for ten 2.5 by 2.5 degree grid cells in western India by driving GLAM with weather data from the multi-model hindcast ensembles produced by the DEMETER project (http://www.ecmwf.int/research/demeter/). The result is a set of crop yield ensembles which contain probabilistic information on crop productivity.

Extreme events such as low rainfall can result in crop failure (defined as yield below a given threshold); thus, the yield ensembles are used to assess the potential for probabilistic forecasting of crop failure. The hindcasts show skill in the prediction of crop failure, with more severe failures being more predictable. Furthermore, ensemble means can be used to predict interannual variability in crop yield. The performance of the multi-model ensemble mean yield is compared to the skill of baseline deterministic simulations using reanalysis data (the European Centre for Medium Range Weather Forecasts forty-year reanalysis, ERA40). The skill of multi-model yield ensemble means is greater than or comparable to the skill using ERA40.

The impact of two key uncertainties, sowing window and spatial scale, is briefly examined. The impact of uncertainty in the sowing window is greater in the ERA40 case than in the multi-model ensemble mean case. Subgrid heterogeneity, however, has more of an impact on the skill of the multi-model ensemble: where there is no skill on the grid scale, there may be skill on the subgrid scale.

The results presented here have implications for yield forecasting on timescales from the seasonal to the multi-decadal (i.e. climate change). These implications are highlighted briefly in the concluding remarks.