Where do the feral oilseed rape populations come from? A large-scale study of their possible origin in a farmland area

Authors


*Correspondence author. Sandrine Pivard. E-mail: sandrine.pivard@polytechnique.org

Summary

  • 1Many cultivated species can escape from fields and colonize seminatural habitats as feral populations. Of these, feral oilseed rape is a widespread feature of field margins and roadside verges. Although considered in several studies, the general processes leading to the escape and persistence of feral oilseed rape are still poorly known. Notably, it remains unclear whether these annuals form transient populations resulting mainly from seed immigration (either from neighbouring fields or during seed transport), or whether they show real ability to persist (either through self-recruitment or seed banks).
  • 2We conducted a 4-year large-scale study of factors involved in the presence of feral oilseed rape populations in a typical open-field area of France. The results were subjected to statistical methods suitable for analysing large data sets, based on a regression approach. We subsequently addressed the relative contribution of the ecological processes identified as being involved in the presence of feral populations.
  • 3Many feral oilseed rape populations resulted from seed immigration from neighbouring fields (about 35–40% of the observed feral populations). Immigration occurred at harvest time rather than at sowing. Around 15% of such populations were attributed to immigration through seed transport.
  • 4The other half resulted from processes of persistence, mainly through persistent seed banks (35–40% of the observed feral populations). This was all the more unexpected because seed banks have not yet been documented on road verges (despite being frequent within fields). Local recruitment was rare, accounting for no more than 10% of the feral populations.
  • 5Synthesis and applications. Understanding the dynamics of feral oilseed rape populations is crucial for evaluating gene flow over an agro-ecosystem. Our results show that, while many feral populations do come from annual seed dispersal, a significant number also result from seeds stored in the soil for several years. In the current context of coexistence and management of transgenic with non-transgenic crops, feral persistence and, especially, the seed bank contribution to the dynamics of feral populations need to be considered seriously. The latter, combined with self-recruitment, indicates a high potential for the persistence of transgenes and the possible emergence of gene-stacking.

Introduction

In agro-ecosystems, gene flow can occur via pollen over space, and via seeds over both space and time. Each crop has its own distinctive characteristics of pollen and seed production, dispersal and potential outcrossing, thus experiencing different levels of gene flow. Moreover, crop and wild plants have lived side by side and exchanged genes for a long time (Squire et al. 2003). Twenty-two of the world's 25 most important food crops can hybridize with wild relatives in some part of their cultivated range (Ellstrand 2003). Understanding and managing gene flow in agro-ecosystems is important, whether for protecting crops from contamination from wild plants or volunteers, for enabling specific production within crops or, more recently, for assessing the possibility of transgene transfer from crops to wild relatives (Snow, Uthus & Culley 2001; Snow et al. 2003; Hails & Morley 2005).

Oilseed rape (Brassica napus L.) is a major crop in western Europe, grown mainly for oil production. It is described as one of the highest-risk crops for gene flow, with many possibilities for gene-stacking, hybridization and introgression (Eastham & Sweet 2002). Crop-to-crop pollen flow is known to occur over long distances (Lavigne et al. 1998; Squire et al. 1999; Rieger et al. 2002; Devaux et al. 2005). Oilseed rape is also cross-compatible with several wild relatives (Chèvre et al. 1997, 2004; Jørgensen et al. 1997). Seed losses from fields are well documented, and oilseed rape shows great ability to establish seed banks within fields and to grow volunteers within other crops (Hails et al. 1997; Pekrun & Lutman 1998; Lutman et al. 2005). A particular characteristic of oilseed rape is its ability to grow on natural and seminatural habitats. These so-called feral populations are widespread on field margins, roadsides or sometimes waste grounds (Crawley & Brown 1995; Charters, Robertson & Squire 1999; Pessel et al. 2001; Ramsay, Thompson & Squire 2003; Crawley & Brown 2004). Not only do these populations provide an interesting case study of fragmented population dynamics, but models show that they could be important for gene-flow management (Colbach, Clermont Dauphin & Meynard 2001; Claessen, Giligan & Van den Bosch 2005; Garnier & Lecomte 2006).

The ecological processes underlying the presence and persistence of feral oilseed rape, in particular the processes related to seed sources, are unclear. On one hand, some studies suggest that oilseed rape behaves as an early successional ruderal, incapable of regenerating in undisturbed habitats (Crawley & Brown 1995; Crawley et al. 2001). Feral populations are then considered as resulting mainly from seed spillage from farm machinery or trucks, or directly from neighbouring fields cultivated the current or previous year (Lutman 2003). In that case, persistence on roadsides would be due to a high rate of local extinctions, compensated, for instance, by recruitment from seeds spilled during transport (Crawley & Brown 1995). On the other hand, there is evidence that the dynamics of these feral populations does not rely only on seed immigration (Squire et al. 1999; Pessel et al. 2001). These studies, contrasting with previous ones, suggest that populations could persist for many years via self-recruitment (defined as the self-replacement of locals by recruitment from seeds produced by residents) or via the contribution of a seed bank.

Thus many questions underlying the presence of feral populations remain unresolved, with huge consequences for gene-flow management. In particular, the relative importance of persistence vs. immigration is crucial: it determines the extent to which these populations can be relays or reservoirs of (trans)genes and contribute to their dispersal and persistence. If feral populations result mainly from seed spillage and thus act only as a seed sink, their contribution will obviously be smaller than if they can persist and become new sources of genes, directly or via a seed bank. Moreover, the longer they can persist, the greater the risk of gene-stacking (Lutman 2003).

We explored the likely origins of feral oilseed rape populations based on an exhaustive 4-year survey at landscape level in a farmland area. Our aim was to understand the factors controlling the presence of feral populations, concentrating on the processes related to seed input, as described above. (i) We evaluated the effect of potential local seed sources as neighbouring fields (imported seeds) and feral populations (self-produced seeds) and the nature of local dispersal (immediate germination vs. delayed germination from the seed bank). Using a regression approach, we modelled the probability of the presence of feral populations on roadsides in the last year of the survey as a function of variables designed to assess underlying processes of presence, and especially to distinguish between seed immigration and persistence sensu lato. (ii) We also evaluated the proportion of feral populations resulting not from local processes, but from seed dispersal at longer distance, considering sites at which no seed would have been imported or produced locally for several years. (iii) We finally addressed the relative importance of the ecological processes identified as contributing to the presence and persistence of feral oilseed rape populations.

Materials and methods

study area and survey

The study area is a 42-km2 production area of winter oilseed rape, centred in the village of Selommes (Loir-et-Cher, central France) and on a silo, to which most of the farmers owning fields within the area take their crops.

Twice yearly, from 2000 to 2003, we made a census of 110 km of roads. The first was made in early April during oilseed rape flowering; the second in early July before harvesting. The census was conducted from a van moving at a maximum of 15 km h−1. Roads were classified into three types (Fig. 1): tracks (42% of roads), one-lane (29%) or two-lane (29%) paved roads. Weed management of roadsides differed depending on road type. Along tracks, field margins are weeded by farmers, while the village or district is responsible for management of paved roadsides. All but one of the two-lane roads led to the silo.

Figure 1.

Distribution of winter oilseed rape fields cultivated within the 42-km2 area from 2000 to 2003 (grey polygons) and of feral populations observed in 2003 (black points) along the road network considered in the study. The latter is composed of tracks (thin grey lines); one-lane paved roads (thin double lines); and two-lane roads (thick double lines), which are often directed towards the silo (black triangle). Cross-hatched polygons represent villages; spotted polygons denote forests. The map was drawn using arcgis ver. 9·1 from the GPS data collected for the study.

At each census, we monitored GPS co-ordinates of all oilseed rape fields and feral plants at every phenological stage from rosette to seed-producing plants. Data were recorded in a database created with the PostgreSQL Database Management System (ver. 7·2) and mapped using arcgis (ver. 9·1). Feral plants were not recorded individually, but in the same set of GPS co-ordinates we considered all plants separated by <10 m. A feral population is thus defined here geographically. In a few cases, the limit between a rape field and its verge was not clear, and it was difficult to determine the extremities of a potential feral population. In such cases we took co-ordinates of the field only and indicated (since 2001) that these fields were bordered by feral plants.

Each year between 2000 and 2003 the area contained between 80 and 100 winter oilseed rape fields, 5·6 ha on average (Fig. 1), occupying 11–14% of the 42 km2 study area. They also bordered on roughly the same proportion of the road verges. The other main crops of this production basin were wheat, barley, maize, sunflower, pea and millet. To our knowledge, no spring oilseed rape was cultivated in the area.

Inter-field margins in Selommes were non-existent. Feral populations were thus restricted to field margins and road verges along the road network. They were dominant features of the road verges, occupying 10–14% of their length across years and comprising several tens of thousands of individuals each flowering season (for example, in 2003 around 22 000 plants were estimated as present in the whole area).

creating the data set

Principle

At this point, the data collected in Selommes were inadequate for understanding the origin of feral oilseed rape populations, notably when only a small part of the feral population matched spatially with a past oilseed rape field, or when different sources were possible. To take into account the spatial distribution of observations and, in particular, to study the effect of local sources on the presence of feral populations, we divided the space into small units. Using a regression approach, we could then explain the presence of feral populations on these small spatial units using available information on potential local seed sources, such as the past presence of oilseed rape fields and feral population in a neighbourhood (defined to take into account the natural seed shadow), and information on environmental conditions. We could also study what would happen when no local source was possible for 4 years, attributing populations to long(er)-distance dispersal or to older seed banks.

Variable creation

As feral populations were restricted to the road verges, the road network was used as the reference system. Roads were divided into 3-m-long oriented segments (distinguishing the two road sides), defining a one-dimensional spatial system composed of 74 002 segments, considered as the spatial units. Co-ordinates of feral populations and fields were projected on the corresponding orientated segments, taking into account the correct roadside.

Several variables with values on each segment were defined. The response variable, the presence of feral population in 2003, was defined for each segment as a binary variable, scoring 1 if there was at least one feral population observed in 2003 during flowering or harvest time. Four ‘permanent’ features (e.g. independent of the year of observation) were also attributed to segments: road number; road type; vicinity of a junction; and village proximity. Road number labelled each of the 101 roads identified in the area. Road type qualified three types of road: track, or one- or two-lane paved road. Vicinity of a junction was a binary variable set to 1 if the segment was <10 m from a junction (6% of segments). Village proximity was also a binary variable set to 1 if the segment was in a village (2·5% of segments).

To take into account the natural seed shadow of fields or feral populations, we then defined a neighbourhood for each segment, composed of the segment itself and the two neighbouring segments on the same roadside. Fourteen variables were designed to describe the history of each segment neighbourhood. (i) Three variables stood for the presence of adjacent feral populations in the neighbourhood of the segment in 2002, 2001 and 2000. These were binary variables set to 1 if a feral population was observed in the segment neighbourhood. (ii) Similarly, three others described the presence of a facing feral population (in the neighbourhood of the facing segment on the opposite side of the road). (iii) Four variables stood for the presence of an adjacent rape field in 2003, 2002, 2001 and 2000. (iv) Four variables represented the presence of a facing rape field. These were all binary variables, except those describing the presence of an adjacent field in 2001 and 2002, for which our data allowed us to set a third value (= 2) when the field was bordered by feral plants for which we had been unable to take precise co-ordinates.

Variable meaning(s) in terms of processes

The 14 history variables were designed to investigate the local seed processes (at a scale of a few metres) contributing to population presence. Depending on the type and year of observation of the potential local sources, feral populations could result from directly imported seeds from neighbouring fields, from self-produced seeds by residents, or from seeds initially stored in a seed bank. For winter oilseed rape, seed immigration from fields occurs at either sowing the previous autumn or harvest at the beginning of the previous summer. It was thus investigated through the presence of neighbouring fields in 2003 and 2002, respectively. The seed bank was scrutinized mainly through the presence of neighbouring fields and feral populations before 2002, while direct local recruitment could be possible only in the presence of an adjacent feral population in 2002.

Longer-distance dispersal could not be investigated in the same way. No variable could be easily connected to seed transport. We thus used the four ‘permanent’ variables for investigating a potential effect of seed transport by assuming that seed spillage frequency and intensity would differ according to the different variable values. Nevertheless, they could not be disconnected from other environmental conditions such as weed management.

statistical methods

Statistical analyses were performed using a mixed-effects logistic model (Pinheiro & Bates 2000) on grouped data (the frequency of segments where the presence of feral population in 2003 equals 1 over the total number of segments for each of the 4534 combinations of factor levels). We modelled the probability of the presence of feral populations on roadsides in 2003 as a function of the set of explanatory variables describing the history of surrounding areas in terms of previous oilseed rape populations, and taking into account some of the environmental conditions. This model was used both to analyse classically the effect of the explanatory variables on the presence of populations (anova), and to predict the presence of populations in situations allowing a better understanding of the relative importance of the variables and of the underlying processes. All the analyses were handled using r software (ver. 2·2).

A bagging algorithm as a precursor to the model

Considering all the variables in the model leading to convergence failures in the estimation process, we decided not to include in the model the less important variables according to a data-mining method called ‘bagging procedure’ (Breiman 1996) and based on regression trees (see Appendix S1 in Supplementary Material for more details on the method and results). Regression tree analysis is an appropriate way to explore large and complex multivariate data sets, as it is free from assumptions about the data distribution and about the relationship between the predictor variables and the response variable, but it is not a good technique for robust parameter estimation or rigorous hypothesis testing. Here we used it as a precursor to the mixed-effects logistic model to determine which variables could be neglected in the final model construction. The least important variables (Fig. S1), which we subsequently eliminated, were the three variables designating the presence of feral populations on the facing roadside and the two variables describing the presence of a nearby village or a junction.

Model construction and analysis

We introduced into the mixed-effects logistic model the 12 remaining variables up to the second-order interaction terms as fixed effects: 11 variables directly connected with local seed sources (presence of adjacent and facing rape fields from 2000 to 2003; presence of adjacent feral population from 2000 to 2002) and the road type connected with longer-distance seed dispersal and environmental conditions. For the same reason, the road number (Rdnb) was also introduced, as a random variable effect to take into account unknown sources of spatial heterogeneity over the region. It was assumed to follow a centred Gaussian distribution with a constant variance. We handled extra binomial variance that could result from dependence between segments or unconsidered sources of variability using an overdispersion parameter, estimated to 2·55. The residual analysis validated the statistical model and, in particular, suggested no modification in the model (Appendix S2). We performed marginal anova to understand the effects of variables.

using the model to evaluate the relative importance of factors and underlying processes

To compare the relative importance of factors in determining the presence and dynamics of feral populations, we then used the model in a predictive way. We first predicted the effect of each single factor under the condition that all other factors were not implicated: set to zero to better separate and compare the strength of underlying processes contributing to the presence of feral populations by comparing probabilities on segments where a single local source of oilseed rape was possible. For this, we used pairwise comparisons followed by Bonferroni correction. In addition, we also predicted the probability of the presence of feral populations for different combinations of fields and feral population presence/absence in the past, for instance when accumulating years of presence of a feral population.

The logistic model fit and the related calculations were handled by the glmmPQL function of the r software.

relative contributions of processes to feral population presence

Relative contribution of long-distance dispersal

Although circumstantial evidence from the analyses suggested that some populations should result from long-distance seed immigration, no model variable was sufficiently direct to evaluate its contribution. We thus considered a subset of segments, called ‘virgin’ segments, which had been neither in the neighbourhood of oilseed rape fields from 2000 to 2003 nor in the neighbourhood of a feral population from 2000 to 2002. On these segments, the presence of a feral population in 2003 was possible only via long-distance dispersal, probably during transport or from a very old seed bank (constituted before 2000). We studied the spatial pattern of virgin segments that were either occupied by a feral population in 2003, or not occupied. From their frequency in the area, some clues about the relative weight of long-distance dispersal and very old seed bank for explaining the presence of feral populations on them and assumptions about the remaining area (see Results), we deduced a rough evaluation of the potential contribution of seed transport to feral populations.

Relative contributions of all candidate processes, including local and long-distance dispersal

We then evaluated the contribution of groups of variables that had been significant in the model to the presence of feral populations in 2003. We considered the following related events: presence of a field (adjacent or facing) in the neighbourhood in 2003, related to local seed immigration at sowing; presence of a field (adjacent or facing) in 2002, related to local seed immigration at harvest; presence of an adjacent population in 2002, essential condition to self-recruitment; presence of a field (adjacent or facing) before 2002 (thus in 2001 and 2000); and presence of an adjacent population before 2002, related to a seed bank role. For each event, E, we calculated its conditional probability given the presence of a feral population in 2003, using Bayes's rule:

P(E/Pop03 = 1) = [P(Pop03 = 1/E)P(E)]/P(Pop03 = 1)

The probability conditional to E was estimated with the logistic regression model described above. The ratio of the probabilities P(E)/P(Pop03 = 1) was calculated directly from the observations using empirical frequencies. The sum of the conditional probabilities exceeded 1 because we put no constraint on variables that are not exclusive. Dividing the probability by this sum gives us a rough evaluation of the relative contribution of these groups of variables and thus potential local underlying processes, taking into account the contribution of transport evaluated previously.

Results

which factors best explain the presence of a feral population in a given place?

Results of the anova for the reduced model are given in Table 1. The presence of an adjacent field in 2002 was clearly the best factor for explaining the presence of a feral population in 2003. Among other significant main effects, we found the presence of adjacent fields and feral populations in 2000 and 2001, the presence of facing fields in 2002, and of marginally facing fields in 2000. As expected, the presence of these past oilseed rape fields or feral populations increased the probability of finding a feral population in 2003, suggesting that both local seed immigration from neighbouring fields and local seed bank were important seed sources for observed feral populations.

Table 1. anova of the logistic mixed model
Parameter*DdlFP
  • Only interactions significant with P < 0·001 are shown.

  • *

    Pop02, Pop01, Pop00: presence of feral populations in the neighbourhood of the segment in 2002, 2001 and 2000; Field03, Field02, Field01, Field00: presence of a rape field in the neighbourhood of the segment in 2003, 2002, 2001 and 2000; FieldF03, FieldF02, FieldF01, FieldF00: presence of a rape field in the neighbourhood of the facing segment; Rdtype: type of the road. See Materials and methods for more detail.

Field022106·17<0·0001
Field001 25·08<0·0001
Pop001 22·03<0·0001
FieldF021 17·35<0·0001
Pop011 13·090·0003
Field012  5·890·003
FieldF001  4·960·026
Field031  3·490·062
FieldF011  2·960·085
Rdtype2  2·300·106
FieldF031  1·610·205
Pop021  0·480·486
Rdtype:Field032 32·30<0·0001
FieldF02:FieldF001 26·73<0·0001
Field03:FieldF021 22·44<0·0001
FieldF03:Field012 21·99<0·0001
Field03:Field001 20·70<0·0001
Rdtype:FieldF022 16·85<0·0001
Rdtype:Pop022 11·26<0·0001
Rdtype:Field014  6·91<0·0001
Rdtype:Field024  6·39<0·0001
Field03:FieldF011 15·980·0001
Field03:Field022  9·430·0001
Rdtype:Pop002  8·220·0003
FieldF03:FieldF001 12·790·0004
FieldF03:Field022  7·220·0007
Field03:FieldF031 11·270·0008
FieldF02:Field012  7·190·0008

Seventeen interactions were significant (P < 0·001), including many interactions between road type and the presence of oilseed rape, including the presence of feral populations in 2002. Quite often, more feral populations in 2003 were estimated on one-lane or two-lane roads than was explained by additive effects, which could be due to seed transport or environmental conditions such as weed management. We also found significant interactions between different factors linked to field presence, while no interaction between factors linked to the past presence of feral populations was significant. At that point, it was noticeable that, through either their main effect or interactions, each introduced variable was significant and thus no possible origin of spatial and temporal seed dispersal could be eliminated for explaining the presence of feral oilseed rape populations.

further clues on longer-distance dispersal in seed transport

In addition to interactions between road type and the presence of oilseed rape, a further indication was given by the significant random road number effect, estimated to 1·36 with a 95% confidence interval of [1·13, 1·63]. This demonstrated that conditions were more homogeneous within a road than between two distinct roads. The random road-effect estimates plotted on the road network (Fig. 2) allowed us to consider spatially which roads had a higher probability of feral plant presence than was predicted by the fixed effects (black lines), and conversely which roads had a lower probability (thick grey lines). Of the six main roads leading to the silo (see Fig. 1), five had positive to strongly positive effects, strongly suggesting seed dispersal during transport.

Figure 2.

Map of the random road effect. Roads were divided into four classes according to the quartiles of random effect estimates: thick black lines, road effect in [1·21, 3·78]; thin black lines, road effect in [0·5, 1·21]; thin grey lines, road effect in [–0·71, 0·5]; thick grey lines, road effect in [–2·1, –0·71].

how strong are the effects of the various factors connected with seed sources?

On virgin segments, in the absence of any possible local source of seeds from 2000, the predicted probability of presence of a population in 2003 was 0·019 (Fig. 3), giving a kind of ‘background noise’ including long-distance dispersal and very old seed banks. The presence of an adjacent rape field only in 2002 increased this probability significantly more than any other variable, suggesting that seed immigration from local fields at harvest is a more important seed source than any other processes. Furthermore, any presence of rape in 2000 (adjacent or facing field and neighbouring population), as well as adjacent feral populations in 2002, also increased the probability of a feral population in 2003 over that of the virgin segments (Fig. 3), highlighting both persistence processes, seed banks and self-recruitment. Probabilities predicted with other factors were not significantly different from that predicted for virgin segments, all road types confounded (these became significant when distinguishing the road type, data not shown).

Figure 3.

Estimated probabilities of the presence of a feral population in 2003 for each predictor variable when all other predictor variables = 0, as well as for virgin segments. Error bars, 95% CI for marginal means of fixed effects in the probability scale. Letters A–D split the variables into four groups according to a multiple comparison procedure. Using the Bonferroni correction, each comparison test between two variables is of a level <0·05/66 = 8·10 – 4. See Table 1 for a detailed description of the variable names.

Accumulating years of the presence of a feral population significantly increased the estimated probability of its future presence, whether (Fig. 4b) or not (Fig. 4a) they followed a rape field. The probability of feral population presence following three consecutive years of feral population presence was not significantly different from the probability estimated in the presence of an adjacent field in 2002 alone (P = 0·41).

Figure 4.

Comparison of probabilities of the presence of a feral population in 2003. Probabilities for segments (a) with feral populations since 2002, 2001 and 2000; (b) with fields in 2000 and 2001 in the absence and presence of following feral populations. Error bars, 95% CI for marginal means of fixed effects in the probability scale.

what is the contribution of long-distance dispersal to feral populations?

Virgin segments contained 6% of the feral populations observed in 2003, on one-third of the study area. Assuming that (a) we knew roughly the proportion of feral populations observed on virgin segments resulting from long-distance dispersal (over old seed bank); (b) most road verges were a possible habitat for feral populations, whether or not they concerned virgin segments; and (c) seed spillage occurred in the same way on the two-thirds remaining area, then we could deduce a rough evaluation of the potential contribution of seed dispersal to feral populations. From partial information available on fields for the period before 2000, 1% of the feral populations in 2003 could match with available field co-ordinates. The 5% remaining, sparsely distributed across the whole area, were thus likely to result from long-distance spillage. Then about 15% of the feral populations in 2003 would probably result from seed dispersal, mainly transport, including some that seemed to come from fields or past populations.

what are the relative contributions of all the highlighted processes?

Assuming this result, we then calculated the relative contribution of some groups of variables to the presence of feral populations in 2003, using Bayes's formula. The presence of a field (adjacent or facing) in 2002 contributed as much as the presence of a field in any of the two previous years, and would each explain 20–25% of the feral populations present in 2003. The presence of adjacent populations before 2002, as well as the presence of a field in 2003, would explain about 15%, as much as that evaluated from spillage during transport. Lastly, the presence of an adjacent population in 2002 would explain about 10% of the feral population present in 2003 (Fig. 5).

Figure 5.

Relative contribution of groups of variables to feral populations observed in 2003, including the transport contribution evaluated through the study of virgin sites.

Discussion

Quantifying the relative importance of different processes involved in the presence of feral oilseed rape populations is crucial to gene flow assessment at the landscape level, because of the potential role of feral populations as reservoirs or relays of genes (Colbach, Clermont Dauphin & Meynard 2001). It is necessary to address these issues on an ecological timescale and with respect to the landscape context of the populations observed (Sork et al. 1999). In studying a cultivated species, we had to cope with several particularities: oilseed rape is introduced massively to the area each year, voluntarily as a crop and inadvertently as a feral, which creates a very different situation from classical studies of wild population dynamics. Moreover, feral populations have a very particular one-dimensional distribution, whether along road verges or field margins. We thus developed a method to handle the data surveyed and then used a regression approach to predict the probability of presence of feral populations in 2003 as a function of variables describing the history and landscape context of the corresponding sites. Variables were chosen to assess the underlying processes of presence, and especially to try to distinguish as far as possible between seed immigration (from neighbouring rape fields, other populations and longer-distance dispersal, mainly from transport) and persistence sensu lato, either from local recruitment or germination from the seed bank. We investigated which related factors explain the distribution of feral oilseed rape populations, then predicted the presence of population in situations allowing us to understand better the relative importance of the factors and of the underlying processes, and finally to address their relative importance.

spatial scale of analyses

Projection on 3-m segments proved very valuable for studying the origin of feral populations. As for short-distance and temporal dispersion, explaining only a few metres at once was necessary to discriminate the possible sources contributing to the presence of feral populations. Indeed, a population forming a homogeneous continuum of plants could be a mixture in terms of origin and genetic pool. We chose a priori a segment length in the order of magnitude of (i) natural seed shedding (≈0.5 m in Colbach, Clermont Dauphin & Meynard 2001); (ii) the GPS precision (less than 1 m); and (iii) the road width (a few metres), justifying the separation of the roadsides. With hindsight, the choice was validated by the huge difference in the effects and contributions of fields and populations, depending on whether they were adjacent (at <3–4 m) or on the other roadside (up to 10 m).

survey reliability

We estimated detection error in the feral population censuses by comparing the data presented here with an independent demographic survey carried out by searching six roadsides (10% of the road network) on foot during the same period (unpublished data). Of the 118 populations found using this method of intensive search, we had detected all but eight small populations totalling 18 plants (10 rosettes, eight flowering plants, of which only one produced seeds in June). All these plants occupied about 50 m on the 20 km of censused roads. The measurement error was thus only about 0·25%.

processes involved in the presence of feral populations

Local seed immigration from neighbouring fields

The results showed unambiguously that imported seeds from adjacent fields at harvest time were the first purveyors of feral populations. This was not surprising, as overall losses at harvest are large in winter oilseed rape, representing ≈10% of harvest (several thousand seeds per m2), of which about half results from natural shedding (Price et al. 1996; Hobson & Bruce 2002; Gulden, Shirtliffe & Thomas 2003). A smaller proportion of populations appeared to result from seed losses at sowing in the same year. Moreover, the effect of fields adjacent to the target segment was always larger than that of fields across the road: either seed projections from harvesting machinery seldom crossed roads, or feral populations might be founded mainly by seeds falling from mature pods of plants growing on the field margin, consistent with results of Price et al. (1996). Overall, immigration from neighbouring rape fields was an important cause of feral population presence, likely to be responsible for 35–40% of the feral populations.

Longer-distance immigration

New populations can be founded by long-distance seed dispersal, mainly via human-related vectors such as farm machinery or during transport. There is evidence for seed losses during transport: many feral oilseed rape populations in the UK have been shown to result from seeds falling from trucks (Crawley & Brown 1995; Charters, Robertson & Squire 1999). Oilseed rape seeds were also observed far from any field or population under tunnels in Germany (Von der Lippe 2004). In our study, the lack of a direct variable concerning transport prevented us from directly assessing its contribution to feral population presence. Significant variables were impossible to separate entirely from other human activities or environmental effects, such as weed management schemes differing between tracks and paved roads. Some clues were apparent, however, such as highly significant interactions between population presence and road type in the logistic model. The clearest indication was obtained from the random road number effect: roads directed towards the silo had large positive random effects, harbouring more populations than predicted from the fixed-effects variables, which has no reason to be explained by weed management. Studying a subset of virgin segments on which populations could be founded only by long-distance immigration or a very old seed bank (from before 2000) was crucial in providing conclusive evidence for the foundation of populations by dispersed seeds at long distance, and for providing a rough contribution of 15% of the feral populations observed in 2003, under assumptions. The assumption that most road verges were a possible habitat for feral populations should be largely correct, except for some small parts of roads located in villages and forests. The assumption that seed spillage was similar on all roads was more questionable: considering the model results and, in particular, the random road effect, it might lead to a slight underestimate of the contribution of long-distance dispersal. This contribution should thus be considered as indicative.

Persistence

Accumulating years of the presence of a feral population significantly increased the estimated probability of its future presence. In particular, three consecutive years becomes equivalent to the presence of a field the previous year. However, information on the presence of feral plants over time does not in itself reveal the processes involved in population persistence (Charters, Robertson & Squire 1999). Contrary to the past presence of fields, the positive effect of which can be explained in only one way, various processes can explain apparently persisting populations. These include local recruitment, but also persisting seed banks or even frequent seed spillage over particularly favourable habitats. It was decisive here to assess and compare the effect of each variable alone.

Local recruitment

In the study area, about 15% of populations observed at flowering also produced seeds in 2002, allowing for local recruitment (data not shown). Our results showed that the single presence of a feral population in 2002 significantly increased the probability of finding a feral population in 2003 over that on virgin segments, strongly suggesting an important role of local recruitment. Nonetheless, the relative contribution of adjacent feral populations present in 2002 revealed that local production of feral individuals cannot be responsible for more than 10% of the 2003 feral populations. This value, calculated over the whole data set, should be considered as a maximum because it did not exclude alternative explanations such as seed banks.

Seed banks

Persisting oilseed rape seed banks are frequent within fields for at least 5 years (Hails et al. 1997; Pekrun & Lutman 1998; Lutman et al. 2005), but have not yet been documented on road verges where the soil is more compacted and seed banks are thus less likely. Nevertheless, some studies suggest that seed banks contribute to the persistence of feral populations (Charters, Robertson & Squire 1999; Pessel et al. 2001). Here we obtained good evidence for persisting seed banks from the positive effect of all types of oilseed rape presence before 2002 (main effect and interactions), confirmed unequivocally by the significant positive effect of any presence of rape in 2000 (field or adjacent population) when the other factors were set to zero. The difference between effects of years 2000 and 2001 might be due to large-scale climatic effect, with the year 2000 being more favourable to seed bank constitution or preservation (Crawley & Brown 2004). About 40% of feral populations in 2003 may have grown from a seed bank, as much as those from seed losses from fields. Thus feral populations should definitively not be considered extinct even if plants fail to appear in any one year (Charters, Robertson & Squire 1999), as has already been demonstrated in natural populations (Eriksson 1996; Stocklin & Fischer 1999).

Conclusion

The dynamics of feral oilseed rape populations appears to be an important determinant of gene flow over an agro-ecosystem (Claessen, Giligan & Van den Bosch 2005). Globally, our results indicate that about half the feral populations in the study area were founded by seeds that fell the previous year at harvest or during sowing. The other half result mostly from seeds that remained in the soil for more than 1 year or, in smaller proportions, from seeds produced within the feral population the previous year. This large contribution of a seed bank to the dynamics of feral populations should be taken into account in models of gene flow at the scale of an agro-ecosystem, such as genesys (Colbach, Clermont Dauphin & Meynard 2001). Indeed, because of this persistent seed bank, plants from cultivars grown in different years could produce hybrids between cultivars and might allow transgene-stacking if GM crops were to be grown. Further, plants from cultivars that are withdrawn from the market will be found in feral populations and may contribute pollen to nearby fields.

Acknowledgements

We are very grateful to E. K. Klein for statistical advice, A. Garnier, J. Shykoff and G. Squire. We also warmly thank the editors and two anonymous referees for their helpful comments and corrections on the manuscript. This project was financially supported by the French Ministry of Research and Technology.

Ancillary