Keywords

1 Introduction

Because of environmental and health issues and the resulting changes in agricultural policies, weed management strategies must be rethought from scratch to rely little or not at all on herbicides. The switch from a single highly efficacious technique, that is, herbicides, to a combination of partially efficient preventive and curative techniques (Liebman and Dyck 1993; Liebman and Gallandt 1997) needs models to explore the vast range of possible combinations, to assess long-term effects and the many services (e.g. trophic resources for pollinators and pest enemies) and disservices (e.g. competition for soil resources, host for crop pests) depending on weeds.

In order to understand and predict the variability in effects observed for the different cropping techniques in a large range of situations without reparameterisation, mechanistic models are best. Such models decompose the life cycle of weeds and crops into elementary processes depending on biophysical effects of cropping systems, in interaction with biophysical variables (Colbach et al. 2005; Colbach 2010). Indeed, it is not sufficient to quantify the average effects of techniques; farmers also need to know the probability of success of a given management strategy and the risk of obtaining the opposite effect of the one they were originally aiming at (Colbach et al. 2014a).

For that purpose, the required model needs to consider most of the cropping system components, even if they do not directly target weeds as do herbicides or mechanical weeding. Indeed, any effect on the crop or the environmental conditions will also affect weeds. Model inputs must also include pedoclimatic conditions to take account of regional differences and, most importantly, to integrate interactions between cropping systems and environmental conditions (Colbach 2010). As weed seed banks persist for several years in the soil (Lewis 1973), a comprehensive model must allow for simulations over several years or even decades to assess how today’s decisions could affect weed flora and crop production during the years to come (Colbach et al. 2014a). This model needs a daily time step to be consistent with the temporal scale of farming operations and the interactions with pedoclimate. The model should also be multispecies, both in terms of weeds and crops. Indeed, arable fields include several dozens or even hundreds of different weed species (Fried et al. 2008), and crop diversification is an important lever of integrated weed management (Liebman and Dyck 1993; Liebman and Gallandt 1997).

This chapter will present such a model and describe how it is used to design agroecological weed management strategies. FlorSys (Colbach et al. 2014b, c, 2017c; Gardarin et al. 2012; Mézière et al. 2015; Munier-Jolain et al. 2013, 2014) is a “virtual field” (in silico) approach which allows for the simulation of weed and crop growth and reproduction on a daily basis over the years, on which cropping systems can be experimented and a large range of crop, weed and environmental measures estimated (Fig. 11.1).

Fig. 11.1
figure 1

General representation of the (1) research model FlorSys (Colbach et al. 2014c; Gardarin et al. 2012; Mézière et al. 2015; Munier-Jolain et al. 2013) which simulates crop growth and weed dynamics from cropping system, weather and soil inputs based on a mechanistic representation of biophysical processes at a daily time step (3D representation), and the (2) metamodel DeciFlorSys which directly estimates weed services and disservices from cropping system inputs (Colas et al. 2020) (Nathalie Colbach © 2018)

Full size image

2 FlorSys: The “Virtual Field” Model

2.1 Input Variables

The input variables of FlorSys (Fig. 11.1) consist of (1) a description of the simulated field (i.e. daily weather, latitude and soil characteristics); (2) all the simulated crops and management operations in the field (including dates, tools and options); and (3) the initial weed seed bank size and composition which is either measured on soil samples or, more feasible, estimated from regional flora assessments (Colbach et al. 2016).

2.2 Weed and Crop Life Cycle

The input variables influence the annual life cycle which applies to both annual weeds and crops, with a daily time step (Fig. 11.1). Pre-emergent stages (e.g. surviving, dormant and germinating seeds, emerging seedlings) are driven by soil structure, temperature and water potential. The crop–weed canopy is represented in 3D with an individual though simplified representation of each crop and weed plant. Post-emergent processes (e.g. photosynthesis, respiration, growth, etiolation) are driven by light availability and air temperature. At plant maturity, weed seeds are added to the soil seed bank; crop seeds are harvested to determine crop yield. FlorSys (Colbach et al. 2014c; Gardarin et al. 2012; Mézière et al. 2015; Munier-Jolain et al. 2013) is currently parameterised for 25 annual weed species (including different populations differing in terms of herbicide resistance) and 33 cash and cover crop species typical of temperate European agroecosystems.

2.3 Effect of Cultural Practices

Life cycle processes also depend on the dates, options and tools of management practices (tillage, sowing, herbicides, mechanical weeding, mowing, harvesting), in interaction with weather and soil conditions on the day the operations are carried out. For instance, weed plant survival probabilities are calculated deterministically depending on (1) management operations (tillage, herbicides, mechanical weeding, mowing, harvesting) and their options (e.g. tillage depth, tool, speed), (2) biophysical environment (e.g. soil moisture, canopy density) as well as (3) plant morphology and stage. The actual survival of each plant is determined stochastically by comparing this probability to a random probability. Survival after herbicide spraying also depends on plant genotype.

2.4 Parameterising Many Contrasting Species

A mechanistic approach is important to ensure that a model allows continuously synthesising knowledge (Colbach 2010) but it requires an enormous amount of parameters, which hinders the addition of new species to the model. This is the reason why Gardarin et al. (2012, 2016) developed a new methodology based on functional relationships to estimate difficult-to-measure model parameters from easily measured species traits, databases and/or expert knowledge. The validity of this approach was checked on weed species, for the critical emergence stage (Gardarin 2008) and at a multiannual scale (Colbach et al. 2016; Pointurier et al. submitted to Ecological Modelling).

2.5 Assessing Crop Production and Weed (Dis)Services

To simplify the comparison of cropping systems and to make simulations more accessible to policy makers, crop advisors and farmers, the detailed outputs are translated into indicators assessing crop production, as well as weed-borne agroecological services and disservices. FlorSys production indicators comprise crop yield in terms of weight and energy (Fig. 11.2). Indicators of weed disservices describe weed harmfulness for crop production and were developed in cooperation with farmers and crop advisors (Colas et al. 2020; Mézière et al. 2015). Direct (crop yield loss and harvest pollution by weed debris) and indirect weed harmfulness (weed-borne pests) affecting crop yield, as well as technical (harvesting problems due to weeds blocking the combine) and sociological harmfulness (weed field infestation as a proxy of the farmer’s worry of being thought incompetent by his peers, even if there is no actual effect over yield loss) were included.

Fig. 11.2
figure 2

Simplified representation of spatio-temporal stages and processes in the FlorSys model (Colbach et al. 2014c; Gardarin et al. 2012; Munier-Jolain et al. 2013). (a) Temporal representation of annual life cycle of crops and weeds, showing the 1D representation of the soil seed bank. (b) 3D individual-based representation of the crop–weed canopy, focussing on plant–plant competition for light (Nathalie Colbach © 2019)

Full size image

Weed-service indicators were developed with ecologists and agronomists (Colbach et al. 2020; Mézière et al. 2015; Moreau et al. 2020 in press at European Journal of Agronomy) and reflect the contribution that weeds make to biodiversity and the environment. They consider weed plant diversity (richness and evenness), the role of weeds for feeding three major guilds in the agro-ecosystems (pollinators, farm birds and carabids) and for reducing three physical farming impacts on the environment (nitrate leaching, pesticide transfer and soil erosion).

2.6 Domain of Validity

FlorSys was evaluated with independent field data, showing that crop yields, daily weed species densities and, particularly, densities averaged over the years were generally well predicted and ranked Colbach et al. 2016, Pointurier et al. submitted). However, a corrective function was required to keep weeds from flowering during winter in southern France (e.g. below 46°N) (Colbach et al. 2016).

3 Running Virtual Experiments

In this section, different case studies illustrate how FlorSys is used to run virtual experiments at different temporal and spatial scales, aiming not only to control weed but also to promote weed-based services.

3.1 Efficacy Evaluation of a Management Technique

Integrated crop production methods often delay seeding to enhance weed control. For example, delayed sowing in winter crops allows more time for false (or stale) seed bed technique favouring autumnal weeds to emerge during the summer fallow, thus resulting in a reduced weed seed bank at crop sowing and, hopefully, in a lower weed emergence inside the crop (Moss and Clarke 1994). However, this strategy is only efficient if the targeted weed seeds are not dormant before crop sowing. Moreover, its efficiency varies considerably with environmental conditions, mainly with soil moisture. Indeed, false seed bed techniques work best when the targeted weed seeds are moist. Conversely, if the sowing operation is combined with a superficial tillage and carried out when the soil is moist, tillage triggers additional weed germination, resulting in increased weed emergence inside the crop (Fig. 11.3).

Fig. 11.3
figure 3

Effect of the last tillage date associated with the sowing operation (carried out on 26 Sept.; 3 Oct.; 10 Oct.; or 17 Oct.) on the cumulated autumnal emergence of grass weeds (e.g. Alopecurus myosuroides Huds.) in winter wheat in Burgundy simulated with the monospecific prototype of FlorSys. The arrows indicate the sowing date relative to the daily weed emergence (dashed line) in case of the earliest sowing (26 Sept.). Grey areas indicate days where the soil was too dry for germination. Delayed crop sowing allows to avoid the earliest weed emergence flush and reduces in-crop weed emergence (sowing on 3 Oct. and 17 Oct. vs 26 Sept.). If the crop is sown shortly after the soil was remoistened by rain (10 Oct.), the associated tillage triggers a germination flush resulting in a huge increase in weed emergence after sowing. No additional triggering occurs if the soil is tilled in dry conditions (3 Oct.) or in continuously moist soil (17 Oct.) (based on Colbach et al. 2005) (Nathalie Colbach ©)

Full size image

To evaluate the success rate of delayed sowing and the risk of unexpected side effects, a virtual experiment was carried out with FlorSys (Table 11.1). Five wheat sowing dates were tested in two French regions, and each was repeated with ten different weather series. The initial weed seed bank consisted mostly of Alopecurus myosuroides, an autumnal grass weed typical of winter-crop rotations in Eastern France. The frequency analysis of the simulation output showed that delayed sowing indeed decreased weed emergence in crops in both regions, in average. But, in Northern France, sowing had to be delayed until 7 Nov. to avoid all risk of increasing weed emergence. In Burgundy, where the soil often is too dry for germination in early October, a residual risk of increased weed emergence persisted until mid-November (Table 11.1).

Table 11.1 Effect of delayed winter-wheat sowing (combined with a power harrow) on autumnal grass weed emergence (e.g. Alopecurus myosuroides Huds.) in the crop simulated with the monospecific prototype of FlorSys. Probability of occurrence (% years) that weed emergence increases or decreases relatively to the initial sowing date on 3 Oct. (based on Colbach et al. 2014a)
Full size table

3.2 Multicriteria Long-Term Evaluation of Cropping Systems

The main interest in using a model such as FlorSys lies in the long-term and multicriteria assessment of comprehensive cropping systems. Figure 11.4 shows an example of model use in interaction with crop advisors, to assess the advantages of crop diversification, introducing spring crops into the usual 3-year winter rotation consisting of oilseed rape, wheat and barley. The analysis of the yield-loss dynamics over time demonstrated the necessity to evaluate innovations in the long term. For example, spring pea presented the highest yield loss of all tested crops (Fig. 11.3.A), but the yield loss in the following wheat crop was consistently lower than in wheat following oilseed rape or sunflower, even though both these crops presented a much lower yield loss than spring pea. Consequently, in average for the long-term evaluated time horizon, the rotation including spring pea performed much better in terms of crop production and weed harmfulness than the 3-year reference rotation and as good as the 5-year rotation including sunflower and spring barley (Fig. 11.5b).

Fig. 11.4
figure 4

Effect of crop diversification on crop yield loss due to weeds simulated with FlorSys. Annual means averaged over 10 weather repetitions with a Burgundy pedoclimate as a function of time (vertical bars show intra-annual standard-deviation averaged over time) (a) and multicriteria evaluation of weed (dis)services averaged over rotation (b) for winter oilseed rape/winter wheat/winter barley (OWB, red line); winter oilseed rape/winter wheat/spring pea/winter wheat (OWpW, blue line); winter oilseed rape/winter wheat/sunflower/winter wheat/spring barley (OWsWb, green line). Means followed by the same letter are not significantly different at p = 0.05, using a least-significant difference test. (Colbach and Cordeau 2018b; Colbach et al. 2010) (Nathalie Colbach ©). § TFI is treatment frequency index (a herbicide at full dosage over whole field = 1)

Full size image
Fig. 11.5
figure 5

Landscape patterns that were virtually experimented with FlorSys. Small landscapes consisted of four 3-ha fields and a typical pedoclimate from Aquitaine (south-western France). (a) Landsharing scenarios based on a single diverse rotation (soybean/maize/wheat/maize), differing in the number of crops present each year in the landscape; (b) Landsparing scenarios with fields aiming to maximise crop production (“Prod”) but converting part of the field into permanent grass strips (green 10 m wide strips); and (c) Landsparing scenarios with varying proportions of contrasting cropping systems in the landscape, either aiming to maximise biodiversity (“Biodiv”) or crop production (“Prod”) (based on Colbach et al. 2018) (Nathalie Colbach ©)

Full size image

In addition to conventional biodiversity and harmfulness criteria, FlorSys also allowed to assess performance indicators that are almost impossible to evaluate under field conditions, such as weed-based food offer for pollinators or farm birds. In the present example, crop diversification allowed to improve all analysed performance indicators (i.e. increased biodiversity and crop production while reducing weed harmfulness and herbicide use).

This approach is invaluable to assess innovations before they are actually authorised and introduced into cropping systems, for instance to evaluate the impact of genetically modified herbicide-tolerant crops and the accompanying changes in cropping practices on biodiversity (Bürger et al. 2015; Colbach et al. 2017b), herbicide resistance in weeds (Colbach et al. 2017c; Sester et al. 2006) or harvest quality, for instance in terms of fatty acid content (Baux et al. 2011) or genetic impurities (Sausse et al. 2013).

3.3 Upscale to Landscapes

Model-based evaluation is also helpful when upscaling from the field to the landscape level. Switching scales could be necessary when weeds disperse to neighbour fields (via seeds or pollen) as the management of a given field will influence what happens in neighbour fields. Pollen dispersal is an issue if the immigrant genes change the fitness of the native population, which is particularly the case for herbicide resistance. Even without propagule exchange, working at the field cluster or landscape scale can be pertinent when there are trade-offs between crop production and biodiversity conservation. In such a case scenario, models can contribute to decide whether landsharing or landsparing is more adequate (Colbach et al. 2018). Indeed, semi-natural habitats and landscape crop patterns contribute to weed dynamics by locating favourable habitats, both in time and in space (Petit et al. 2013).

FlorSys allows tackling some of these questions by simulating several fields and/or semi-natural habitats in parallel (Colbach et al. 2018). At seed shed, weed seeds as well as shattered crop seeds are dispersed from a source plot to neighbouring habitats. Seed dispersal distance increases with weed plant height and decreases with seed mass; and it is higher for seeds dispersed by animals and wind than for those dispersed by gravity (Colbach et al. 2018; Thomson et al. 2011). The dispersed seeds then colonise new fields and habitats or integrate existing populations, contributing to wild plant biodiversity but also negatively affecting crops.

The spatially explicit model proposed by Colbach et al. (2018) allows to virtually experiment different landscape management scenarios, aiming to reconcile crop production with biodiversity conservation, either at field (landsharing) or landscape (landsparing) levels. Three series of scenarios were simulated over 28 years and 10 weather repetitions, using maize-based cropping systems (Fig. 11.5). The simulations showed that landsparing scenarios were better than landsharing, resulting in high crop production and medium biodiversity at the landscape scale (Table 11.2). Landsharing scenarios always produced less biodiversity and less crop production. The more crops were grown each year in the landscape, the more the weed impact on production and biodiversity increased.

Table 11.2 Effect of landsharing and landsparing scenarios (Fig. 11.5) on weed (dis)service indicators at the landscape scale (Colbach et al. 2018)
Full size table

3.4 Disentangling Effects

Many of the evaluations of the previous sections would also be possible with empirical models. One of the major advantages of process-based models such as FlorSys is their ability to disentangle complex interactions in the agroecosystem, often better than in situ field experiments could do. For instance, yield loss is notoriously difficult to assess in field conditions because it is next to impossible to produce a weed-free control identical to the weed-infested treatments except for the presence of weeds. Similarly, farm-field networks (i.e. a large number of farm fields that are monitored in terms of management practices and, for example, crop yield or weed infestation) are usually inadequate to assess the effect of individual management techniques, even influential ones such as herbicides, because farmers reason each technique depending on other cropping-system components (e.g. mechanical weeding, tillage and rotation) (Colbach and Cordeau 2018b).

FlorSys was used to unravel some of these interactions by monitoring a virtual farm-field network covering contrasting production situations with several hundred cropping systems recorded in seven regions ranging from northern France to northern Spain (Colbach and Cordeau 2018a). The effect of herbicides was discriminated from that of other management practices by comparing the simulated weed floras and yields of the recorded cropping systems to those of these same systems minus herbicides (and without any other changes in practices). Moreover, the authors differentiated the relative effects of weeds and management practices on crop production by comparing the yields of simulations run with and without weeds. Also, management practices effects on weeds were separated from their reciprocals (i.e. effects of weeds on management practices) by simulating the recorded cropping systems without adapting the practices to the simulated weed floras. Long-term weed harmfulness was also assessed by looking at weed dynamics and weed-caused yield loss over succeeding years instead of considering only annual data. As a result, Colbach and Cordeau (2018a) were able to show that (1) weed–crop biomass ratio at crop flowering was the best indicator of the year’s yield loss (Fig. 11.6); (2) herbicide use intensity was not correlated to either weed variables or yield loss, because farmers compensated reduced herbicide use by other preventive (e.g. false seed bed techniques) and curative measures (e.g. mechanical weeding); (3) average weed biomass during crop growth and yield loss increased by +116 and + 62% (averaged over rotation) respectively when herbicides were eliminated without any other change in management practices; and (4) effects were more visible at multiannual (rotation) than the annual scales.

Fig. 11.6
figure 6

Generic function predicting grain yield loss (%, i.e. 100 t t−1) in annual crops from the ratio of weed biomass vs crop biomass at the onset of crop flowering established on a virtual farm-field network simulated with FlorSys. Each data point is 1 year of one cropping system and one weather repetition out of a total of 272 cropping systems × 30 years × 10 weather repetitions. Red line fitted to data with non-linear regression (Nathalie Colbach © 2018) (based on Colbach and Cordeau 2018a)

Full size image

This kind of virtual farm-field network can be used for other purposes, for instance to track innovations among farming practices. FlorSys simulations demonstrated the importance of crop diversity in rotations to control weed harmfulness with few or no herbicides (Colbach and Cordeau 2018a). Three types of strategies could be identified among the investigated farmers’ cropping systems, which differed in terms of rotation, tillage strategy and so on. The strategy based on a summer-crop monoculture relied heavily on herbicides as well as mechanical weeding to limit weed-caused yield loss. Conversely, two other strategies diversified crops, with longer rotations, crop mixtures, cover crops and temporary grassland. Combined with well-reasoned tillage, crop diversification allowed for reducing herbicide use while limiting yield loss.

More generally, these farm-field networks allow to identify which management techniques drive weed (dis)services. The previous studies simulated actual cropping systems practiced by farmers or proposed by crop advisors, but the network can also be extended with randomly constructed systems to run sensitivity analyses. This was actually the approach used when metamodelling FlorSys into a decision-support system (see section 11.4.1) where various statistical methods were used to identify the most influential management techniques. Table 11.3 shows an example of a ranking of management techniques in terms of their effect on weed contribution for protecting the soil from erosion and nitrate leaching. This analysis shows that tillage, particularly deep and/or inverting operations, was the major determinant of weed-based soil protection while rotation and herbicides had much less impact.

Table 11.3 Identification of key management techniques influencing weed (dis)services in a virtual farm-field network consisting of several hundred cropping systems from six French and Spanish pedoclimates simulated with FlorSys over 27 years and 10 weather repetitions (based on Moreau et al. 2020). Example of weed-based protection from soil erosion and nitrate leaching averaged over 27 simulated years, identifying key techniques with LASSO regressions
Full size table

The same simulation approach, combined with statistical methods usual in ecology, such as RLQ or fourth corner, can identify crop and weed traits that drive crop production and weed (dis)services (Colbach et al. 2014d; Colbach et al. 2017a; Colbach et al. 2017b; Colbach et al. 2019). Table 11.4 shows an example where the aim was to identify which weed-morphology parameters that drive weed harmfulness for crop production in average over many contrasting cropping systems and pedoclimates. The most damaging weeds in terms of crop yield loss and harvest pollution were the ones that occupied space earlier and faster, starting with a large leaf area at emergence and with larger and/or thinner leaves (larger SLA). Later in the season, shading neighbours with taller plants per unit biomass (larger SPH) also becomes important, but lateral space occupation is still an issue, as wider heavier plants (larger HPW) with a uniform leaf area distribution (lower MLH) are more damaging. When shaded, the damaging weeds react by shifting their leaves topwards (increase in MLH).

Table 11.4 Which weed plant-morphology parameters drive weed harmfulness for crop production?
Full size table

4 Decision Support with Stakeholders

4.1 DeciFlorSys: A Decision Support Tool

The previous sections show examples of how FlorSys can be used to track innovations in existing farming practices, test, diagnose and fine-tune prospective cropping systems proposed by farmers and crop advisors, produce expert knowledge for policy makers and so on. However, FlorSys still remains a research model inadequate for direct use in participatory workshops as (1) it requires numerous input variables to be assigned and parameters to be tuned, (2) its mechanistic and individual-based approach induces a high algorithmic complexity and very slow simulations and (3) it evaluates cropping-system candidates rather than actually designing these candidates (Colbach 2010). In order to address these limitations, we transformed FlorSys into a Decision Support System (DeciFlorSys). DeciFlorSys gathers three operational tools, each addressing one of the above issues. The three tools all derive from the metamodelling of FlorSys using sensitivity analyses and machine learning, and they were co-designed with future users (Colas et al. 2020). Instead of using detailed inputs, DeciFlorSys uses aggregated inputs corresponding to meta-decision rules at the rotation scale (e.g. proportion of spring crops in rotation, frequency of mouldboard ploughing) (Fig. 11.7). It directly predicts weed (dis)service indicators, without calculating detailed crop and weed variables (Fig. 11.1). The three DeciFlorSys tools are (1) a table showing the cropping system components to be changed as a priority, (2) decision trees showing how to combine management practices to reach a given goal in terms of weed (dis)services and (3) a predictor based on random forests (AI technique) that calculate the performance of the cropping-system prototypes, with a much faster response time than FlorSys and easier to handle than the parent model.

Fig. 11.7
figure 7

Decision tree identifying cropping system rules for reaching a series of 13 weed impact profiles based on a multivariate regression tree linking weed impact indicators simulated by FlorSys to combinations of cropping system variables (blue: rotation, green: sowing/harvest, brown: tillage, black: weed control), based on 4350 cropping systems from 20 French regions. Indicator values averaged over years and weather repetitions were coloured from white (minimum) to green (maximum) for biodiversity, from white to red (maximum) for harmfulness to crop production. Uncoloured cells show standard-error including weather effects and variability among systems in a branch. Cross-validation error = 0.106, for indicator range of variation rescaled to [0,1]. The tree can be read top-down to get an idea of the performance of a proposed combination of management practices, or bottom-up to identify the practices corresponding to a target performance (Floriane Colas © 2019) (based on Colas 2018)

Full size image

While the DeciFlorSys predictor is as good as FlorSys to rank cropping systems, it cannot adequately evaluate effects that strongly interact with pedoclimatic conditions, such as the effect of tillage timing with respect to soil moisture (Colas 2018).

4.2 Use of Models to Promote Integrated Weed Management

Both FlorSys and its derivate, DeciFlorSys , have been used by our research team and crop advisors in participatory workshops with farmers. Implicating farmers in cropping-system design is essential, as innovations proposed by scientists are often disregarded by farmers because they are incompatible with farming constraints (Meynard et al. 2018) or with farmers’ risk perception and management (Wilson et al. 2008). Crop advisors can also be reticent to promote the necessary changes (Pasquier and Angevin 2017). In this context, models are invaluable teaching tools to propagate knowledge and promote innovations via training sessions, participatory workshops and role-playing games (Hossard et al. 2013; Martin et al. 2011; Meylan et al. 2013; Sausse et al. 2013). This is particularly true for easy-to-use models such as DeciFlorSys, which allow stakeholders to directly and immediately see the consequences of changes in their practices in their particular production situation.

Using models with farmers and crop advisors is somewhat different than when using them for research purposes (Fig. 11.8c vs. b). During the workshops, farmers start from their own experience, they are implicated in all steps and get an immediate feedback, all of which makes the resulting solutions more acceptable to them. Conversely, the risk of missing highly performant solutions and remaining inside current conventions is much higher. The best approach for investigating a larger range of possible solutions is automatic optimisation algorithms (Fig. 11.8a) which have already been used with simpler and faster models than FlorSys (Bergez et al. 2010) and are now being adapted to FlorSys (Maillot et al. 2019).

Fig. 11.8
figure 8

Three ways to use the FlorSys tools to design innovative cropping systems with a step-by-step improvement on an initial cropping system S0. (a) Optimisation algorithms manage all steps in interaction with FlorSys, except fixing the aims and constraints for the novel cropping systems, which are determined by a group of experts. (b) When running virtual experiments, experts fix aims and constraints, compare the simulated performance of the systems to these aims and propose innovative systems, from expert knowledge and the decision tree of DeciFlorSys. (c) Participatory workshops including farmers often start with a system that performed badly in the field; innovative systems are proposed by a group of interacting farmers and other experts using a variety of tools and these systems are evaluated by the predictor component of DeciFlorSys to benefit from an immediate feedback that sets off another round of system design (Nathalie Colbach © 2019)

Full size image

5 Discussion and Conclusion

FlorSys is one of the very few process-based models that include all the key mechanisms that are relevant for cropping system and does this at a sufficiently precise scale to produce realistic results. To overcome the trade-off between process analysis and decision aid (Colbach 2010), the detailed simulated outputs were aggregated into indicators to support decisions (Bockstaller et al. 2008), and the knowledge synthesised in the mechanistic research model was further extracted and summarised as the empirical (meta)model DeciFlorSys which is easier to use. This dual approach allowed us to synthesise knowledge on the functioning and effects of crop diversification at different levels of detail and make it available to different stakeholders, consisting of scientists, crop advisors, farmers and policy makers. It is also essential to continue including new knowledge, by adding new crop and weed species as well as management techniques.

All these advantages are subject to the model’s prediction quality, which must be confirmed by comparing model simulations to independent observations or expertise (model evaluation). This step is even more crucial for a mechanistic model aggregating data and models from different teams and disciplines, to make sure that the new entity produces consistent results. Though this step has been carried out for FlorSys, it also pointed to a major drawback of complex mechanistic models (i.e. the difficulty to find adequate data for evaluating the model and its many submodels).

The possibility of continuous model evolution is crucial as, despite its complexity, FlorSys (and its derivate DeciFlorSys ) neglect several processes that are essential for the more innovative cropping systems, particularly in a context of crop diversification, input reduction and climate change. For instance, including competition for soil nitrogen would improve the assessment of legumes or drought-resistant crops in rotations and mixtures.

The synthesis of the various case studies demonstrated the usefulness of FlorSys not only for synthesizing knowledge on biophysical processes implicated in cropping system effects but above all for producing emerging knowledge on the functioning of the agroecosystem, and for promoting this knowledge among farmers. In terms of integrated weed management, the many studies carried out with FlorSys to date demonstrated that, generally, (1) weed damage can be controlled with few or no herbicides if the cropping system is consistently redesigned, (2) many conclusions in terms of crop diversification only have a local validity, (3) which proves the need for flexible rules and (4) the usefulness of models such as FlorSys and optimisation algorithms to establish these rules.