The suggestion to incorporate rusticity traits in breeding strategies as a way to temporarily or durably 'hard-code' variability management should not be seen as a pessimistic attempt to downplay potential tactical uses of seasonal forecasts in Sudano-Sahelian agriculture or (even worse!) underrate the critical importance of agroclimatic risk analyses. Rather, we believe that before forecasting skill and smallholder endowment improve, there is room for a parallel and renewed effort in the agrometeorological early estimation of crop production. A much larger array of data sources (from climate models and remote sensing), finer understanding of crop growth and development (from process-based models) and stochastic data assimilation (DA) techniques now allow a more operational 'invigoration' of probabilistic agroclimatology by looking at 'weather within climate' (Hansen et al. 2005) in the context of facilitating investment in rainfed agriculture (Cooper et al. 2005). Predictability at intermediate intra-seasonal (-20-60 day) timescales has been somehow neglected in favor of more fashionable seasonal products, but holds promise in the short term as it will benefit from enhanced representation of continental forcings (Céron 2004) and ongoing projects such as AMMA (2005). Experimental hybrid monthly forecasts are routinely published by ECMWF since October 2004 with the objective to bridge the gap between extended weather and seasonal timescales, a priority focus area from a climate science perspective (Grassl 2005). Figure 19.4 proposes a schematic procedure to improve final model yield estimates using such in-season rainfall forecasts and bi-weekly satellite biomass observations (from ASTER) in a sequential DA framework. Sequential DA is computationally more efficient than variational DA recently tested for crop yield estimation (Guérif and Duke 2000; Bach and Mauser 2003). It can accommodate a wider range of uncertainty as Monte Carlo ensemble generation allows for any statistical form of time/space correlation in error structure (Crow and Wood 2003), and can propagate the full probabilistic climate information into yield estimates along with measurement and model uncertainties (Jones et al. 2006). It is better suited for near-real time applications oriented towards the prediction of future system states that is key to early warning systems.
Was this article helpful?