What do crop models say?
And what don’t….
Long time has passed since I posted my last post, and now we approach to a new Conference of Parties (CoP) in Durban. But this is not about the conference, not at all. This is about how I see the agricultural research going in the context of climate change: models, models and models.
Projections of future climate are the sole means we (as agricultural researchers) have to deliver some information about the impacts of climate change on crop production. Climate projections are input into crop models (mathematical approximations of the physiological and phenological mechanisms of crop species), which themselves yield projections, this time not of future climates, but of future agricultural yields. These are then interpreted, most of the times correctly. Nevertheless, sometimes projections are taken as certain futures and, I’ve read it some times: some state this or that will happen by 2050 (or whichever future year). This is truly not true. Projections of climate change are designed to provide a range of plausible futures that can be used to evaluate the responses of different systems so that decisions (related to adaptation and mitigation, mostly) can be made with a scientific base.
This misunderstanding arises from the massive gap between the agricultural modelling and climate modelling communities. A large proportion of this gap arises from the difference in scales at which both communities operate (or can operate). Agricultural researchers (often) use plot-scale crop models, whereas climate researchers normally provide data at large-scales (typically in the order of hundreds of km, with some exceptions). Matching these spatial scales is a tricky issue.
Two possible solutions exist: (1) increase climate model resolution, or (2) design large area crop models. The former (at the time writing) is unfeasible due to lack of computing capacity, so here I will hold on the latter. Various process-based models exist that describe processes at large scales, from which probably the most famous is GLAM. GLAM was specifically designed to operate at large scales, in contrast to other models such as those in DSSAT, that operate at plot-scales. Hence, one can assume that both models are incompatible. So I ask, are they really incompatible?
So, I took large-scale data (yields, soils, rainfall, temperature and radiation) for the state of Gujarat in India (the square marked with “GJ” in the India map). The data had previously been used for calibrating GLAM, so I tried to model the response of groundnut to large-scale environmental conditions using CROPGRO-PNUT (cv. TMV-2, same cultivar as for GLAM), a plot-scale crop model. No surprise: it did not fit well the data, however, it captured the interannual variability fairly well. So I decided to limit yields using a soil fertility factor, and voilá: see what I got.
Both models performed reasonably well, with a root mean square error of 281 (GLAM) and 227 kg/ha (CROPGRO) over the whole period. Correlations were 0.74 and 0.84 for GLAM and CROPGRO, respectively, and the yield gap parameter (YGP) of GLAM was 1.0, whereas the SLPF (soil fertility factor) of CROPGRO was 0.8, not that different, huh?
Further, both models represented the inter-annual variability of crop yield, as it can be seen in the figure (left). Therefore, the performance of both models combined (ensemble) seems to be a good proxy for yield variability. Quantifying this agricultural model uncertainty is probably a subject of study in the Agricultural Model Intercomparison and Improvement Project (AgMIP). So they’d be rather happy when reading this.
So I conclude that it is highly important to (1) use process-based models (even if they’re simple in formulation) so that we ensure we give our analyses a statistical, yet biological sense, (2) not rely in one single approach and quantify the relevant uncertainties in the modelling process, (3) ground our analyses with data, (4) use the relevant approaches at the appropriate scales, and (5) not take our predictions as absolute truth. This way, we will surely improve our understanding of modelling methods and climate change impact assessment science, hence decreasing the risk of mal-adaptation.
Acknowledgment Thanks to Andy Challinor and Jim Watson, from University of Leeds for the data and ideas.