Global food and development policies are increasingly being supported by crop models, but current modelling approaches are unfit for this purpose. The models in use, many of which were developed in the 1970s and 1980s for high-input monoculture systems, ignore critically important aspects of sustainable agriculture. And they do not work for complex agricultural landscapes common in developing countries, systems in which trees are integrated with crops, and which, in addition to crop yields, bring numerous benefits to people and the environment. A shift towards more holistic crop modelling is urgently needed. If crop models continue to neglect systems and outcomes that cannot be modeled with precision, their use will keep directing us towards purely yield-optimizing systems that perform poorly with regard to critical development objectives such as ecological sustainability and social fairness.
Special Think Piece, first published on ICRAF's blog here.
I recently attended the first major meeting of the global crop modelling community in a long time — the International Crop Modelling Symposium held in Berlin. Most major players in the discipline were there, including some of the people who laid the foundations for all the big models in the 1970s and 1980s. Consequently, many of the presentations discussed the history of crop modelling and the impressive progress that the field has made since its early days. Models originally designed to simulate crop growth at the plot level are now used for large-scale food security assessments and for policy advice up to the global level. The ability of models to simulate crop responses to weather, soil and management has greatly improved and colleagues from all around the world presented a number of very promising results.
Gaps in current crop models
From an agroforestry perspective, however, the state of the modelling art is quite sobering, especially when it comes to smallholder systems in many developing countries. Almost all convincing applications of crop models focus on simple systems. In most cases, they consider only pure stands of crops grown in monoculture. Simulations of intercropping are rare, and at the conference I didn’t come across any presentations that reported on process-based models that include trees in agricultural landscapes.
Even some of the major constraints encountered by farmers around the world are currently not well represented. These include the impacts of pests, weeds, diseases and labour constraints, but also water-related phenomena such as waterlogging and extreme droughts. These are very important problems affecting the productivity of many farms, and they are not considered in the models most crop modelers rely on.
Another omission I noticed is that essentially all model applications I saw at the conference focused exclusively on projecting yields. Agroforestry, however, like many other types of production systems, produces a range of other ecosystem services. These include tangible products such as food from trees, fuelwood, medicine, soil fertility and construction materials, and also less visible services such as water cycle regulation, wildlife and pollinator habitat, and cultural identity for people living in farming landscapes.
I realize that adding all these complicated features to models will be very challenging. But is it acceptable to base policy advice on models that ignore them? Let’s imagine what advice can arise from such models.
If the only inputs my crop model considers are attributes of the crop variety, soil conditions, management and weather, only certain types of recommendations can emerge from them. Basically, these are restricted to advice on crop and seed selection, crop development, and on-farm management measures such as fertilization and irrigation. Such models cannot show the need for better pest control, the importance of labour peaks at certain times of the year, the need for land tenure security, or the benefits from better linking farmers to markets. In short, we should not expect them to deliver advice that accounts for the complex situations many farmers find themselves in.
Similarly, if the outputs of our model are restricted to crop yields and ignore other ecosystem services, the advice that their use leads to may end up being rather one-sided. It will place great emphasis on food production but likely fail to consider other objectives of rural development, such as sustainability, resilience, employment, soil conservation and nutrition. Indeed, failure to factor in long-term impacts and off-site effects of interventions has resulted in the unsustainability of many of today’s farming systems. We can observe many negative consequences of this one-sided view, including the pollution of freshwater and ocean systems caused by over-fertilization in mono-cropping systems, or the ecologically and often economically impoverished landscapes that dominate many of the world’s ‘most productive’ agricultural regions.
The biggest problem in using models that strongly simplify complex systems is that system features they fail to include may implicitly be valued at zero. But few agricultural experts – including farmers – would agree that biotic stresses (pests, diseases and weeds) have no effects, or that ecosystem services other than yields have no value. So the values that models assign to these processes or services – zero – is most certainly wrong.
If all cropping systems produced about the same quantity of ecosystem services, this wouldn’t matter, because inclusion of these services wouldn’t affect our preference for one type of system over another. But there are clear differences. Many agroforestry systems, for instance, produce large amounts of fuelwood, fruits and other foods, fodder and construction materials. They often improve soil fertility, regulate water cycles, reduce wind speeds, conserve soils, sequester carbon and provide shade. Models that do not take into consideration all these contributions systematically undervalue agroforestry systems – and many other system types as well. Instead, they tend to favour high-input monoculture systems that produce high yields but not much else – systems that may even undermine the natural resource base they rely on.
The limitations of current crop modelling approaches
Crop modelers may – and sometimes do – argue that as the discipline progresses, the current gaps in the models will be filled. But is this a realistic expectation?
The ambition so far has been to simulate all important ecological processes that are related to crop growth in a way that allows fairly precise prediction of crop yields. For simple systems, this has been somewhat successful, but can we hope to understand ecological interactions in more complex systems to a comparable degree? And if we can, will such models have huge data requirements that severely restrict their application in practice?
I find it hard to imagine a convincing process-based model operating in the way common crop models do that includes the array of constraints many farmers face and the many dimensions of agricultural production that holistic, sustainability-oriented advice to development decision-makers should consider. Such a model would be massively complex and rely on so many input parameters, many of which would be almost impossible to obtain, that it would be essentially unusable.
The challenges that crop modelers face also have to do with the unpredictable nature of many drivers that affect farms. For example, crop losses due to agricultural pests are very difficult to predict because pest life cycles can be complex. They can depend on processes at landscape or even coarser scales that crop models don’t normally consider and that have not been studied sufficiently. Models that aim to make precise deterministic predictions cannot adequately account for phenomena for which precise predictions aren’t feasible.
In short, modelling strategies relying on precise representation of all relevant processes may be doomed to fail for all but the simplest agricultural systems, and they are unlikely to provide us with comprehensive evaluations of the total value of agricultural impacts that go far beyond simple yield predictions.
A fresh view on crop modelling
If there is little hope of making precise performance models of complex agricultural systems, can we not model them at all? Well, we may have to think of a different strategy for getting this done.
First of all, we may have to accept that precise representation of all important processes is not within reach. This is a serious problem for the way many crop models operate, but it can be overcome if we adopt probabilistic modelling approaches.
Probabilistic models do not require precise inputs. They can already give us answers, if we manage to describe how much we know about a variable of interest. We can provide such descriptions in the form of confidence intervals (ranges, in which we’re sure that we’ll find the true values) or probability distributions. Probabilistic models can work with such inputs. Naturally, if we aren’t precise about model inputs, we cannot expect precise outputs. But probabilistic models can also express outputs as probability distributions, which describe the plausible ranges of outcomes that we can expect.
If we don’t know much about our systems, such models can only give us rather vague answers, but ones that may help us considerably in improving policy and management decisions. These answers would reflect the limited knowledge that we have, which arguably is preferable to very precise answers relying on numbers in which we have very little confidence.
Abandoning precision in crop modelling opens up a number of new opportunities. We may, for example, question the precision with which common crop models describe many of the ecological processes. This precision becomes apparent only when we look under the hood of the models, but it may be worth asking to what extent models rely on best-bet assumptions and empirical relationships that are difficult to defend and may not be universally valid.
If we don’t have to be precise, we have the new option of including in our models things to which we can’t assign definite numbers. Going back to the ecosystem services that agroforestry systems provide, we normally can’t pinpoint a precise value for fuelwood production. But we may be able to say that a hectare of a certain tree-based farming system produces anywhere between 1 and 5 tonnes of fuelwood per year. This range then allows us to include fuelwood in production value calculations, and it keeps us from neglecting this important product just because we cannot predict its production with certainty.
We can proceed in a similar manner to value other ecosystem services, in order to arrive at a robust range of the Total Economic Value that agricultural systems provide. Basing land-use advice on such ranges would raise the chance that our recommendations not only promote high-yielding cropping systems but also favour equitable and sustainable production practices.
Holistic crop modelling is more realistic crop modelling
A shift towards a more holistic approach to crop modelling that does justice to the complexity of the modelling challenge, the data limitations we encounter in many places and the unpredictability of many processes could be easier than it may seem. Techniques to address this challenge exist, albeit outside the array of methods that most crop modelers are familiar with.
I’ve been working for some time with risk assessment and decision analysis methods, which are designed to support risky decisions on complex systems, about which we have limited knowledge. They address challenges that have much in common with the quest to predict the performance of complex cropping systems for which we are equally far from knowing everything we would like to know. These methods are widely used in many other fields, such as software risk management, oil exploration and public health planning, but not commonly in crop modelling.
One of the main principles of decision analysis is that we should consider all factors that seem important for the system we work on. We should integrate everything we know about the system, including data and expert knowledge, but honestly acknowledge the limitations of our knowledge. And we should fully consider our uncertainty rather than basing evaluations on best-bet assumptions. If, after doing this, we are not satisfied with the precision our model provides, we can use procedures known as Value of Informationanalysis to identify the key uncertainties in our model, which we can then address through targeted research. Key techniques of decision analysis are Monte Carlo simulations and Bayesian Network modelling, both of which hold great promise for crop modelling.
Such approaches would be a radical departure from the way crop modelling is commonly done today. They would allow us to make performance predictions that are more holistic and as a result more realistic, and expand the range of systems that we can hope to model in the foreseeable future. We would no longer be restricted to yield predictions for very simple systems, but build the capacity to address the concerns of a wider array of farmers and to consider not only crop production but also the host of other ecosystem services that sustainability-oriented land use advice should include. As a result, we can expect wiser policy and management decisions on intervention choices.