METHODOLOGICAL INSIGHTS: The role of satellite image-processing for national-scale estimates of gene flow from genetically modified crops: rapeseed in the UK as a model


Luisa J. Elliott, NERC–Environmental Systems Science Centre, The University of Reading, Earley Gate, Whiteknights, Reading RG6 6AL, UK (fax +44 118 378 6413; e-mail


  • 1There is concern over the possibility of unwanted environmental change following transgene movement from genetically modified (GM) rapeseed Brassica napus to its wild and weedy relatives.
  • 2The aim of this research was to develop a remote sensing-assisted methodology to help quantify gene flow from crops to their wild relatives over wide areas. Emphasis was placed on locating sites of sympatry, where the frequency of gene flow is likely to be highest, and on measuring the size of rapeseed fields to allow spatially explicit modelling of wind-mediated pollen-dispersal patterns.
  • 3Remote sensing was used as a tool to locate rapeseed fields, and a variety of image-processing techniques was adopted to facilitate the compilation of a spatially explicit profile of sympatry between the crop and Brassica rapa.
  • 4Classified satellite images containing rapeseed fields were first used to infer the spatial relationship between donor rapeseed fields and recipient riverside B. rapa populations. Such images also have utility for improving the efficiency of ground surveys by identifying probable sites of sympatry. The same data were then also used for the calculation of mean field size.
  • 5This paper forms a companion paper to Wilkinson et al. (2003), in which these elements were combined to produce a spatially explicit profile of hybrid formation over the UK. The current paper demonstrates the value of remote sensing and image processing for large-scale studies of gene flow, and describes a generic method that could be applied to a variety of crops in many countries.
  • 6Synthesis and applications. The decision to approve or prevent the release of a GM cultivar is made at a national rather than regional level. It is highly desirable that data relating to the decision-making process are collected at the same scale, rather than relying on extrapolation from smaller experiments designed at the plot, field or even regional scale. It would be extremely difficult and labour intensive to attempt to carry out such large-scale investigations without the use of remote-sensing technology. This study used rapeseed in the UK as a model to demonstrate the value of remote sensing in assembling empirical information at a national level.


The introduction of genetically modified (GM) crops into the UK or elsewhere poses a spectrum of possible risks, as well as benefits, that relate to both the consumers of the crops and the environment. There has been particular interest in the possibility that gene flow from GM crops to their wild or crop relatives will result in the formation of transgenic hybrids, and ultimately will cause unwanted environmental change (Arnaud et al. 2003; Wilkinson et al. 2003). Within the UK there are at least 16 commercially important crops that have wild relatives with which the crop is able to hybridize spontaneously (Raybould & Gray 1993). Of the GM crops currently under field trial in the UK, rapeseed Brassica napus L. provides the greatest scope for interspecific transgene movement; there are nine cross-compatible recipient species identified, with Brassica rapa L. ranking as the most likely to form hybrids (Scheffler & Dale 1994). There are several factors to consider when evaluating whether hybrids can form between two related species in the wild. These include the physical proximity of the two species, the synchronicity of flowering, the presence and strength of interspecific barriers and the pollen- and seed-dispersal capabilities of the crop (Glover 2002). In this instance, natural hybrids between rapeseed and B. rapa have already been reported in the wild (Stace 1991) and so the feasibility of hybridization has been demonstrated empirically. Nevertheless, the qualitative appearance of a single hybrid does not inevitably lead to unwanted environmental change. To a large extent, this is clearly dependent upon the identity of the transgene, but it is also partly reliant on chance and the abundance and distribution of hybrids. Moreover, rare and isolated partly sterile hybrids may be unlikely to cause widespread transgene recruitment within the recipient species, whereas common and broadly dispersed fertile hybrids could very rapidly lead to the ubiquitous presence of the transgene. At the same time, the feasibility of employing risk management measures to prevent hybridization relies on the efficacy of the method proposed (e.g. placement of the transgene on the chloroplast genome; Daniell 1999), the number of hybrids within a legislative region and the predictability of where hybrids will occur. Knowledge of the distribution and abundance of hybrids is an important element needed for any model aiming to predict the rate of transgene spread at a national scale. Likewise, knowledge of regions where hybrids are most frequent is a vital component for any post-release monitoring of GM crops. Prior to this study, however, there have been no estimates of hybrid distribution or abundance for any GM crop–recipient combination to cover any country in its entirety.

When attempting to calculate hybrid number in a nation, it is important to differentiate between local hybridization, where crop and recipient grow together, and distant hybridization events, mediated by the long-range dispersal of pollen. In both cases, it is imperative to be able to describe the distribution of both crop and recipient to high resolution. This can be achieved most effectively with the aid of remote sensing.

Brassica rapa can be subdivided into three types, wild, weedy and ruderal. The wild type grows almost exclusively on riverbanks, the weedy form grows within broadleaf crop fields (typically rapeseed) and the ruderal is an inhabitant of disturbed land such as building sites. The focus of this research was on the former two ecotypes because the ruderals only occur occasionally in generally short-lived populations. The two ecotypes studies differ with respect to the likelihood of hybridization and also in the scope and nature of any resultant ecological change. High rates of transgene recruitment into the weedy B. rapa are predicted due to the intimate contact between crop and recipient (Jorgensen & Anderson 1994). While the presence of such hybrids may exacerbate an existing weed problem, there is only limited scope for ecological repercussions in the wider (non-agricultural) environment. On the other hand, low to modest numbers of hybrids are anticipated in riverside B. rapa populations (Wilkinson et al. 2000) but the scope for ecological change is considerably greater. Several earlier studies that have attempted to predict hybridization between these species were based on small-scale field experiments (Manasse 1992; Scheffler, Parkinson & Dale 1995; Downey 1999). However, there are innate difficulties in attempting to extrapolate from the results of such work in order to predict hybrid frequencies if GM varieties were grown on a commercial scale. For example, a study using the Cambridge Environmental Research Consultants’ Atmospheric Dispersion Modelling System (CERC-ADMS; Hunt et al. 2001; to model rapeseed pollen dispersal predicted a concentration of 1% of the maximum level at 100 m downwind of a 10-m square field, but empirical data showed that for very large fields this level occurs around 3 km (Hunt et al. 2001). This demonstrates that hybridization rates are affected by the size and shape of the fields. Another study looked at the mean dispersal distance of rapeseed pollen by insect vectors (Manasse 1992). It was found that an increase in the separation distance between patches led to an increase in the mean dispersal distance, and a lower plant density led to more pollinator movement. These results indicate that gene-flow rates are affected by the spatial orientation and density of the fields, suggesting that observed gene-flow patterns will be highly dependant on the design of experimental plots. Hence, it is more useful to adopt empirical approaches to measure hybridization, using commercial-sized fields and natural populations of wild relatives, to parameterize predictive models.

In common with other nuclear genes, transgenes are generally inherited in a Mendelian fashion, and so it is appropriate to study gene flow among non-GM populations to model transgene movement following commercial release of GM cultivars (Gliddon 1994). Difficulty lies in identifying, locating and quantifying potentially rare hybrid plants at a landscape or even national scale. This study set out a series of ways in which satellite remote sensing could be used to assist in hybrid localization for the evaluation of gene flow from a crop at this scale. We used the detection of hybridization between commercial (non-GM) rapeseed and wild and weedy B. rapa in the UK to illustrate the potential of the strategies proposed.

Materials and methods

satellite image acquisition and classification

The characteristic coloration of a rapeseed crop during its flowering period (vivid yellow at visible wavelengths) can be exploited by remote-sensing technology to identify fields of the crop over large areas. The set of fields thus identified may then be used in a number of ways. As a precursor to discussing these, we first outline the rationale behind the acquisition of such satellite images and their subsequent classification.

Landsat thematic mapper (TM) satellite images ( were chosen as the main data type for use in this project for several reasons. First, the Landsat programme has been running since 1972 and continues to collect multispectral digital data of the earth's surface from space. Hence it is a good source of both current and historical data sets. The pixel size of 30 m provides sufficient resolution to locate rapeseed fields, but not the populations of wild relatives. As each complete satellite image covers 185 × 170 km of the earth's surface, it is possible to study whole countries with relative ease and at low cost. Images taken by the Indian remote sensing satellite IRS-1C ( are also suitable as they are spectrally and spatially similar to Landsat TM images.

The area planted with autumn-sown rapeseed over the entire UK (expressed as the proportion of rapeseed in each 2 km2 of land) was obtained for three consecutive years (1995–97) from the EDINA national data centre (hosted by Edinburgh University Data Library, Edinburgh, UK). The mean number of hectares of rapeseed per 2-km square was calculated across all years and plotted to reveal regions containing the highest density of the crop. Landsat TM scenes were acquired to provide maximum coincidence of rapeseed cultivation (based on this distribution) on the one hand and B. rapa populations (Perring & Walters 1990) on the other. In this way, satellite coverage was provided of the areas in which sympatry between the crop and wild relative was most likely (Fig. 1). A total of 8·25 Landsat TM scenes, plus one IRS scene, was acquired for this purpose. The images were collected on the Landsat-5 TM, Landsat-7 TM and IRS-1C LISS-III satellite sensors during the flowering season of rapeseed (May) in various years from 1990 onwards. Low cloud cover was employed as the primary criterion on which to select year–region combinations for image acquisition.

Figure 1.

Mean hectares of winter-sown rapeseed grown in the UK per annum per 2-km square. Black areas, 0·0; green areas, 0·1–5·0; purple areas, 5·1–10·0; red areas, 10·1–15·0; dark orange areas, 15·1–20·0; light orange areas, 20·1–25·0; yellow areas, 25·1–30·0; white areas, 30·1 plus; grey areas, Republic of Ireland (not studied). Overlaid onto this map are the approximate outlines of the Landsat TM and IRS images acquired for this research. Pink square, TM North (1992); green square, TM Lincolnshire (1992, 1995, 2001); dark blue square, TM Oxfordshire (1990, 1995, 2001); turquoise square, IRS Kent (1998); red square, TM Kent (1998, 0·25 scene).

Fields of rapeseed were identified in the images using either an unsupervised classification technique known as ISODATA (Ball & Hall 1967) or a supervised method called maximum likelihood. In order to perform classification, a ‘training’ set of pixels comprising known rapeseed pixels (see below) was compiled for each image. This procedure needed to be carried out separately for each image because the spectral properties of vegetation can vary between images, due to various factors including flowering stage, arrangement of leaves on the plant, nature of the background (e.g. soil type and moisture content) and the viewing and illumination angles (Mather 1999). A mask was created for each image, with the identified rapeseed pixels included and all other ground types excluded. These masked images were then used for further analysis.

Training data sets were typically based on knowledge from ground reference observations. However, national agricultural statistics are confidential at this level of resolution, making historical ground reference data difficult to obtain for the UK. Thus, it was only possible to obtain a limited amount of ground reference information for 3·25 of the images. An alternative method for compiling the training sets for the remaining images was therefore required. The characteristic yellow appearance of flowering rapeseed means that the spectral signature of the crop is distinct from other land cover types during the month of May. For example, we acquired a 0·25 scene of the Kent area taken in May 2003 and obtained ground reference data for in excess of 120 rapeseed and non-rapeseed fields. Results showed that there is a very low false negative rate (1·5% of the known rapeseed pixels were not collected during the masking procedure) and a zero false positive rate (0% of the non-rapeseed pixels were incorrectly classified as rapeseed). This indicates that flowering rapeseed can easily, and accurately, be segregated from the other land cover types in these (and other) Landsat TM images.

When a false colour composite of a TM image is viewed using bands 4 (0·76–0·90 µm), 5 (1·55–1·75 µm) and 3 (0·63–0·69 µm) to represent red (R), green (G) and blue (B), respectively, the rapeseed fields can be seen as pink (Fig. 2). It is also possible to view the rapeseed in a true colour composite, in which RGB is set to the visible red, green and blue wavelengths (bands 3, 2 and 1), showing the rapeseed fields clearly as yellow (Fig. 2). However, if the rapeseed is not in full flower, the near infra-red and middle infra-red bands (bands 4 and 5) seem to be important for distinguishing the rapeseed from other green land cover types, and hence it is more appropriate to view the TM images in false colour (Fig. 2). For the IRS image, using bands 3 (0·77–0·86 µm), 4 (1·55–1·70 µm) and 2 (0·62–0·68 µm) for RGB presented a false colour composite similar to those obtained from the TM images (Davenport et al. 2000). Accordingly, it should be possible to use photo interpretation of the false colour composites to assemble the training sets of rapeseed pixels. Identification of rapeseed clusters, where possible, was carried out using training sets derived from both photo interpretation and ground reference observations, and the accuracy of the former method was tested. The images for which no ground reference observations were available were analysed using only photo interpretation.

Figure 2.

Rapeseed in real and false colour composites of Landsat TM images. Images (a) and (b) are an example taken from the Oxfordshire 2001 TM image and images (c) and (d) are an example taken from the North 2001 TM image. Images (a) and (c) are true colour composites in which the RGB is set to the visible red, green and blue bands, showing rapeseed fields as yellow (outlined in red). Images (b) and (d) are false colour composites in which RGB is set to the near infra-red, middle infra-red and visible red bands, showing rapeseed as pink. The two images were taken within a day of each other (12 May 2001 and 11 May 2001 for Oxfordshire and North, respectively), however, the flowering stage of the northern fields is likely to be less advanced and hence the colours vary between images.

An accuracy assessment of the photo-interpretation method was based on the results of 3·25 Landsat TM images for which ground reference data were available. For all images, the same pixels were identified to be rapeseed regardless of the classifier method used. It was therefore concluded that photo interpretation provides a suitable method for rapeseed detection in Landsat TM and IRS-1C images of the UK. Consequently, the remaining five TM images and the IRS image were classified using the photo-interpretation method. The spectral curves of all identified rapeseed clusters were analysed to verify their identity. The mean spectral curve for rapeseed was determined by using the radiances of all clusters identified using ground reference observations. The radiances of all clusters identified by photo interpretation were plotted with the mean to check whether their curves followed a similar pattern (Fig. 3), and thus it was possible to assess whether each cluster had been correctly identified as rapeseed. Although the values sometimes strayed outside of the 99% confidence intervals, all of the clusters followed the general pattern of the mean spectral curve and appeared to be rapeseed. The mean spectral curve, together with its confidence intervals, is based only on a small amount of ground reference data and thus it would be unreasonable to assume that the values from all other images should lie exactly within this range. Hence, the masked images were considered suitable for utilization in subsequent steps whether they had been generated using ground-verified or photo-interpreted data.

Figure 3.

The spectral reflectance curve for rapeseed for the 0·45–2·35 µm region (Landsat TM bands 1–5 and 7). The black line shows the mean radiance of rapeseed with the 99% confidence intervals (based on all available ground reference information). The thin grey lines refer to all rapeseed clusters determined using photo interpretation (images: Lincolnshire 1995, Lincolnshire 1992, Oxfordshire 1990, North 2001, 1992). The pale grey bars in the background represent six of the seven bandwidths in which the Landsat TM sensor collects information. Band 6 is excluded from this diagram as it collects information at much longer wavelengths (10·40–12·50 µm). DN= digital number.

wildb. rapab. napushybrids formed by sympatry

By reference to the Centre of Ecology and Hydrology Countryside Survey 2000 ( specimens from herbaria across the UK, local floras, the Botanical Society of the British Isles (BSBI) database (Preston, Pearman & Dines 2002) and direct ground surveys, the positions of wild waterside B. rapa populations across the UK were determined. This information was plotted onto a GTOPO30 digital elevation model (DEM) to determine whether there was any relationship between the growing locations and altitude (this DEM has a 30-arc second grid spacing that equates to approximately 1-km resolution). (PDAAC DEM TOPO30 We found no wild B. rapa at altitudes greater than 155 m above sea level (a.s.l.). To confirm this observation, surveys were carried out that started at the source of a river and ended at lower altitudes. The findings were consistent with the earlier hypothesis and hence all land above 155 m a.s.l. could be excluded as a possible habitat for the wild species. The DEM was resampled to a 2-km resolution to be consistent with the agricultural statistics and all 2-km square areas above the altitudinal limit were said to contain no wild B. rapa.

Given that wild B. rapa grows primarily on the banks of rivers, canals and dykes, digital waterways data were obtained to provide a more accurate description of the location of habitats potentially containing the species. Northern Ireland was excluded on the premise that co-occurrence of rapeseed and wild B. rapa is very unlikely due to a low frequency of both the crop and the wild plants. British waterways were digitized from Ordnance Survey (OS) 1 : 50 000-scale data (CEH Wallingford, Oxon, UK) and represented all watercourses more than 3 m wide. In total, there are 540 000 km of river, canal and dyke banks in Britain, providing a large number of possible habitats for B. rapa. Information collected from ground surveys, British herbariums and local floras was combined to identify the river systems in the UK that contain B. rapa. Collectively, these waterways contain 138 000 km of banks on which B. rapa could grow (117 000 km of riverbanks and 21 000 km of canal and dyke banks). The digital waterways data set was divided according to the presence or absence of B. rapa, and to waterway type, i.e. rivers and canals/dykes. Canals and dykes were differentiated from rivers because B. rapa grows only infrequently along canal and dyke banks but is far more common along riverbanks (Wilkinson et al. 2003).

The spatial relationship between rapeseed cultivation and waterways containing B. rapa was studied to provide a first-level approximation of the degree of co-occurrence between the crop and wild relative. The digital waterways data were converted from vector into raster and overlaid onto the masked satellite images. For all 2-km squares within the satellite images, the number of rapeseed pixels, and consequently hectares, within 30 m of a waterway were counted by carrying out a distance transform from the river pixels (Fig. 4). A 30-m distance was used because this is the width of a TM pixel. A quadratic regression was carried out on these data and the resulting equations were used to predict the hectares of rapeseed growing within 30 m of waterways for all 2-km squares not covered by satellite imagery (using the mean hectares of rapeseed per 2 km as calculated from the EDINA statistics and the waterway lengths per 2-km square). These values were combined with the observed frequency of B. rapa on rivers and canals from direct surveys of 151 km of nine riverbanks and 165 km of four canals (Table 1) and used to predict the number of sympatric B. rapa plants within the rapeseed waterway contact zone. To calculate the number of sympatric B. rapa plants, the hectares of rapeseed growing next to waterways were first converted into lengths of river and canal/dyke banks that this represented. These lengths were then multiplied by the frequency of B. rapa plants per metre of riverbank (0·755) or canal/dyke bank (0·004) for the areas known to contain B. rapa and zero for all other areas. Hence, the numbers of B. rapa plants growing sympatric to rapeseed per 2-km square were predicted and plotted onto a map of the UK (Wilkinson et al. 2003). This map essentially represents the areas where local gene flow between rapeseed and populations of wild B. rapa is possible. The majority of sympatry was predicted to occur in central and eastern England.

Figure 4.

An example of a distance transform between river and rapeseed. (a) Masked satellite image (OSR, rapeseed pixels) overlaid with the river image (R, river pixels). (b) Distance transform from the river. Adjacent pixels in the axial and diagonal distances have distance values of 2 and 3, respectively. A value of 2 represents a pixel within 30 m of a river. In this example, there is one pixel of rapeseed that lies within 30 m of the river (shown in dark grey).

Table 1.  Distances of the various rivers and canals surveyed and the number of Brassica rapa plants observed. One bank of the rivers was surveyed and both banks of the canals
WaterwayDistance surveyedNumber of B. rapa plants
River Thames 67·860 884
River Nene 10·333 936
River Stour  4·310 841
River Avon 39·0 1 741
River Derwent  8·8   374
River Ouse  4·4    48
River Soar  1·1     0
River Trent 10·8 6 200
River Great Ouse  4·5     0
Kennet and Avon canal116·3  1218
Boat trip along Grand Union, Coventry and Ashby canals 49·2     2

Having inferred the location and number of B. rapa plants located within 30 m of rapeseed, the next step was to calculate the hybridization rate in such populations. We measured the abundance of hybrid plants in B. rapa populations during the year following sympatry. This has the advantage of accommodating hybrids formed on the crop and dispersed into the wild population by seed and also provides a measure of surviving hybrid plants rather than of potential hybrids (seeds) that may not survive to reproduce because of reduced fitness. The location of rapeseed fields is subject to continual rotation, however, and sympatry between the two populations can be relatively rare (Wilkinson et al. 2000). Rapid location of sites of sympatry for subsequent study of hybrid frequency was therefore necessary. Remote-sensing images were utilized to reduce significantly the search area for such sites in the field. The images acquired in the year prior to the year of fieldwork were overlaid onto OS maps containing rivers likely to contain B. rapa (e.g. 2001 images were studied for the 2002 fieldwork). Sections of numerous OS maps (1 : 50 000 scale) were obtained from EDINA Digimap ( The masked satellite images were overlaid onto the OS sections and any fields located adjacent to rivers were noted. Sites were selected only if there was public access to the riverbank on which the rapeseed grew in the previous year. Subsequently, the selected sites were checked for B. rapa plants and screened by flow cytometry for putative hybrids on the basis of an intermediate ‘G1 peak’ position between those of diploid B. rapa and tetraploid rapeseed. Hybrid status was confirmed by the presence of an amplification product of appropriate size after polymerase chain reaction (PCR) using the C genome-specific microsatellite marker 83b1 (Wilkinson et al. 2000, 2003). In total, we found 47 hybrids from 3230 plants screened (i.e. 1·46%) across eight sympatric populations (Table 2).

Table 2.  The abundance of hybrid plants detected by flow cytometry and microsatellite analysis in eight riverside populations of Brassica rapa
Population grid referenceRiver bank locationNumber of plants screenedNumber of hybrid plants
454335River Trent  43 0
393244River Avon1186 4
478185River Thames 375 0
489187River Thames 130 1
484262River Nene 380 6
454199River Thames 313 0
452195River Thames 597 0
371166River Avon 20636

Approximately 500 km of riverbanks were analysed in a geographical information system (GIS). Eleven sites were found where rapeseed grew adjacent to a riverbank and public footpath in 2001. These sites were visited in 2002 and B. rapa populations were found at two of the sites. These populations contained four hybrids. Thus, the use of remote sensing meant that it was only necessary to survey 6 km of the 500 km of riverbank studied in order to locate naturally occurring hybrids. In addition, the identification of geographical regions containing the highest densities of riverside rapeseed next to rivers containing B. rapa greatly enhanced the efficiency of foot surveys. The number of hybrids found and the total number of plants that were analysed during these field studies were used to calculate the hybridization rate at sites of sympatry. In turn, this allowed estimation of the number of hybrids forming per annum between rapeseed and sympatric wild B. rapa (Wilkinson et al. 2003).

wildb. rapab. napushybrids formed due to long-range pollination events

We sought to model the effects of long-range gene flow, where crop and recipient are separated by relatively large distances. This required a detailed knowledge of the relative distribution and isolation distances between waterside B. rapa populations and the nearest rapeseed fields. The masked satellite images were overlaid with the pixelated waterways data and the proximity of all waterway pixels to rapeseed was calculated by carrying out a distance transform, this time calculating distances from the rapeseed fields rather than from the waterways. The spatial distribution of the waterways and rapeseed fields was determined to predict the probabilities of a given B. rapa plant having rapeseed growing within specific isolation-distance ranges. We calculated the percentage of waterway pixels with a rapeseed field within the following distance categories: within 30 m; from 31 m to 300 m away; at 300-m intervals until 3 km; and > 3 km. This established the relationship between rapeseed and all stretches of waterway that could contain B. rapa. We made a basic assumption that B. rapa is distributed evenly along all river or canal/dyke pixels. An alternative would be to use the observed distribution of the surveyed waterways to produce a probability profile describing the number of B. rapa plants in each pixel. Application of Monte Carlo principles (Metropolis & Ulam 1949) could be used to extrapolate hypothetical locations of B. rapa populations across the country following the same pattern as in the surveyed sample. In practice, however, providing there is no causal reason why the presence of rapeseed influences the presence of wild B. rapa, on average the two approaches should generate near-identical results. Interestingly, the prediction based on surveyed observations would be subject to error in sampling, through regional variance and, crucially, replication error in performing Monte Carlo runs. Once crop–recipient isolation profiles had been established, it was possible to apply pollen-dispersal functions to predict the number of hybrids that form at each distance (Wilkinson et al. 2003).

The frequency of long-range pollination must be estimated to assess the feasibility of reducing hybrid formation between GM crops and their wild relatives through isolation. Estimates of hybrid numbers forming at various distances from the crop enables the effectiveness of strategies for hybrid reduction to be assessed. Significantly, however, while this approach has practical value for the modelling of initial hybrid formation, models to describe the subsequent spread of genes between B. rapa populations will require a more explicit description of population distributions along the rivers. These can either be extrapolated directly from the existing survey data, under the assumption that inaccessible sites (e.g. minor tributaries) show the same patterns as those sampled, or else a high-resolution (e.g. airborne) remote-sensing approach could be adopted.

weedyb. rapab. napushybrids

In the UK, rapeseed is most commonly grown in a 4-year cycle with cereals (generally wheat and barley). When weedy B. rapa is present, it is generally successfully controlled in cereal crops and so only emerges as a significant weed in rapeseed fields at 4-yearly intervals. In these years, the rapeseed provides an abundance of pollen for the self-incompatible B. rapa plants and a large number of hybrids are produced. Several studies have reported that B. rapaB. napus hybrids generally lack dormancy whereas B. rapa populations exhibit variable levels of dormancy (Landbo & Jorgensen 1997; Linder 1998). If such observations apply generally, then the majority of hybrid seeds will germinate in the year following rapeseed (almost invariably a cereal) and will be killed by the application of broadleaf herbicides. However, a fraction of the hybrid seeds will show genetic- or environmentally induced dormancy, and will emerge in later years. As hybrids are only usually able to flower when rapeseed is grown, variation in the rotation cycle has the capacity to influence profoundly the numbers of surviving hybrids.

We used remote sensing to study variation in the length of the rotation period practised in the UK. Landsat TM images were paired according to the year in which they were acquired. Thus, the overlap region in an Oxfordshire image taken in 1990 and a Lincolnshire image taken in 1992 and two Lincolnshire images from 1992 and 1995 were studied to indicate the proportion of 2-year and 3-year rotations, respectively (Fig. 5). The percentage frequency of the 4-year rotation cycle was inferred from the results for the other two rotational patterns, assuming zero occurrence of a 1-year rotation. The results indicate that 87% of rapeseed fields are grown as part of a 4-year rotation (or longer), 9% in a 3-year rotation and 4% in a 2-year rotation. This approach can be used in conjunction with dormancy rates and seed numbers to calculate the hybridization rate. In the companion paper (Wilkinson et al. 2003) we calculated hybridization rate directly on the basis of hybrid plants observed rather than on predicted numbers of hybrid plants inferred from hybrid seed set, seed dormancy and rotation practice. The empirical approach has the advantage of providing a more reliable estimate of hybrid abundance under current practice but cannot so easily accommodate future changes in rotation practice and is less readily applied to situations elsewhere in the world.

Figure 5.

Rotational patterns of rapeseed (fields shown in white). (a) Rapeseed fields in 1992 in south-west Lincolnshire (1695 fields). (b) Rapeseed fields in 1995 that were also rapeseed in 1992 (137 fields; 8% on a 3-year rotation).

In order to estimate the number of weedy B. rapaB. napus hybrids forming per annum, it is necessary to know the total number of weedy populations within the UK. We performed ground surveys but additionally questioned regional agricultural weed consultants about the incidence and locality of B. rapa weed populations. Reference was also made to the CEH Countryside Survey 2000 and to specimens from herbaria across the UK. The number of weedy populations observed and the total number of rapeseed fields surveyed were recorded and the percentage of rapeseed fields that contained the weed was calculated. Difficulty lies in converting the percentage of infested fields into the total number of fields containing weeds in the UK, because national statistics provide the annual area of cultivation of rapeseed rather than the number of fields. The number of rapeseed fields was calculated in each classified satellite image using a connected component finder, together with the total hectares of rapeseed in the image. The mean field size was then calculated, from which we could deduce the number of infested fields. It is possible that the accuracy of this calculation may be slightly compromised by the farming practice of clustering fields of the same crop to ease harvest, or conversely of splitting fields for economic or experimental reasons (e.g. to minimize the risks of growing a new crop or variety or of adopting an experimental practice). In this study, we adopted the simplistic view that these antagonistic practices will be of minor overall importance and will tend to be mutually compensatory and therefore unlikely to greatly affect our estimate of mean field size. The calculated mean field size was used to convert the annual hectares of rapeseed into field numbers. The total number of weedy populations in the UK was determined by multiplying the percentage of rapeseed fields containing the weed by the total number of rapeseed fields in the UK. The mean field size was calculated as 12 ha. The number of weedy populations was combined with the calculated hybridization rate to predict the total number of weedy B. rapaB. napus hybrids forming per annum (Wilkinson et al. 2003). This calculation has importance for predicting possible agricultural problems caused by the introduction of GM herbicide-tolerant (HT) rapeseed and is useful for determining the most effective strategy for the reduction or avoidance of weedy hybrid formation within GM fields.

Results and discussion

Remotely sensed images can be used to characterize gene flow on a national scale in the following ways.

  • 1To determine the spatial relationship between rapeseed fields and nearby waterways, and thereby to predict the locations of local hybrids in the UK formed when rapeseed and wild B. rapa are in sympatry.
  • 2To search more efficiently for hybrid plants through the use of GIS-assisted mapping of possible sites of rapeseed–B. rapa sympatry. This led to an estimation of hybridization rate and hence the number of hybrids forming in the UK per annum. It could also form a key element of post-release monitoring.
  • 3To determine the distances from each B. rapa plant and the nearest rapeseed field. These data allowed modelling of the amount of rapeseed pollen being received by B. rapa plants, and so the number of hybrids arising from long-range pollination can be calculated.
  • 4To calculate the frequency of the various rapeseed rotation types practised in the UK and hence the percentages of weedy B. rapa–rapeseed hybrid seeds remaining viable by the time the field is next cropped with rapeseed. This provides an alternative to measuring the number of hybrid plants in order to calculate hybridization rates, and can accommodate future changes in land use.
  • 5To establish the mean size of rapeseed fields and hence to convert the national statistics data from hectares into field numbers and the number of weedy B. rapa populations. In combination with hybridization rates, predicted from direct observation of hybrid plants, the number of weedy hybrids forming per annum can be calculated.

The overall aim of this research was to assemble generic data sets that could be used to quantify annual gene-flow rates between rapeseed and B. rapa across the UK and to develop a method that is adaptable for use with other crops and in other countries. If GM crops are released on a commercial basis in the UK, then it is important to quantify exposure to the risks and, if possible, to develop effective strategies for risk management. The value of quantifying hybrid abundance on a national scale lies primarily in setting efficiency targets for any measure that aims to prevent its occurrence. Several strategies to reduce or eliminate hybridization between GM crops and wild relatives have been developed. It is reasonable to expect these approaches to vary dramatically in their ability to depress hybrid abundance. It follows that their utility ultimately depends on F1 hybrid abundance in the legislative region under assessment. This paper describes a method for predicting the number and spatial distribution of F1 hybrids forming annually in the UK.

An important application of this method relates to the risk assessment process. Risk is usually defined as being a function of hazard (an unwanted event leading to harm) and exposure (the likelihood that the hazard will occur). Truly quantitative risk assessment therefore relies on measures of exposure. F1 hybrid formation represents the first of a sequence of exposure events needed for any hazard relating to gene flow from GM crops to wild relatives. This must be followed by introgression and gene spread in order for the gene to become stabilized within a wild population. Overall exposure is represented by the probability of completing every step in the ‘exposure pathway’ and critically depends both on the appearance of hybrids and also their abundance, as subsequent steps do not have a probability of 1 (Wilkinson, Sweet & Poppy 2003). That is, one hybrid is never certain to lead to introgression, spread and eventually to ecological change, irrespective of the enhancement to fitness conferred by a transgene. For these reasons, it is important to quantify F1 hybrid formation in the context of the scale at which regulation is levied, i.e. the national level.

The data generated here have practical utility in identifying those regions in the UK where hybrids are most and least likely to form. This knowledge can be used when sanctioning early field trials to select areas where hybrid formation is lowest, and it will also have a vital role in post-release monitoring. However, it needs to be recognized that although our results suggest that most hybrids form locally, the crop nevertheless has a great capacity for long-range hybrid formation when grown on a large scale. The two ecotypes of B. rapa differ in their exposure for hybrid formation. Wild, waterside B. rapa is generally common, although apparently absent from Scotland, whereas weedy B. rapa is rare and largely confined to the Humber region (Wilkinson et al. 2003). Thus, legislative restriction to prevent cultivation of HT GM cultivars in the Humber region would drastically reduce (but not prevent) the number of GM weedy hybrids formed. It would be far less practical to isolate wild, waterside B. rapa populations as it is unlikely that the restriction of GM cultivation to Scotland would be viewed as either a practical or desirable option. This means that the current practice of assessing the risks posed by transgenic lines under the assumption of gene flow still offers the best approach until measures to repress hybrid formation by at least 10−5 are introduced.

This paper sets out a method of predicting the number of hybrids between rapeseed and B. rapa. Both local and long-range events are studied for the wild (waterside) ecotype in the natural environment, and local hybridization is studied for the weedy ecotype within the agricultural environment. The predicted numbers are presented in the companion paper in the form of mean values with associated standard error terms (Wilkinson et al. 2003). This paper details the methods used for the calculations. At present, we can estimate the number of hybrids per annum only to the nearest order of magnitude. For example, the total number of B. rapa hybrids in the UK per annum is predicted to be 48 000 (± 27 000). This means that tens of thousands of hybrids are expected annually.

The main source of error in the current method is in the distribution of the wild B. rapa. Although extensive surveys have been carried out, our current knowledge is restricted to its distribution along waterways that have a path running alongside. In general, these are the main rivers and we infer that the distribution on the tributaries of these rivers is the same as that of the main river. However, we currently have no information regarding the distribution of the species on the tributaries, which form a substantial proportion of the total river length. Additional data from airborne remote-sensing surveys will be used to refine the current predictions of hybrid locations and numbers. Resolution can be as low as 1 m and thus should detect wild waterside populations, which typically may be 1 m wide × 5–10 m long.

The method outlined here for studying hybrid formation at a national scale could be modified for use with other wild relatives of rapeseed and for other crops. Brassica oleracea is known to be the most likely native species to hybridize with rapeseed in the UK (Scheffler & Dale 1994; Preston, Pearman & Dines 2002). This species grows within the UK at a few known coastal locations (Mitchell & Richards 1979). Thus Landsat TM images could be analysed to locate any sites of sympatry between rapeseed and these known locations, or airborne remote-sensing data could be acquired to locate both the B. oleracea populations and any sympatric rapeseed fields. However, a limitation of using this method for other species is that not all crops can be as easily discriminated in satellite images as rapeseed.

There are four crops that are likely to dominate the GM market in the UK: rapeseed, maize, sugar beet and potato. Of these, rapeseed and sugar beet are the only two with close wild relatives in the UK. Sugar beet can be successfully identified using Landsat TM imagery (Fuller & Parsell 1989) and thus it should be possible to quantify gene flow between sugar beet and sea beet in the UK using similar remote-sensing techniques to those described here for rapeseed. As Landsat satellites cover the globe between 81 degrees north and south of the equator, this method could also be utilized for other countries situated within these latitudes. Furthermore, it could be adapted for other crops that have wild relative species growing within these countries.

Crops such as wheat, barley and potatoes can be identified in Landsat TM images, although multitemporal images are required that take advantage of changes in the crops throughout their life cycle. In the UK, images taken between mid-February and mid-April can be used to distinguish cereals and grassland from root crops; images from the first 2 weeks in July are required to discriminate barley from wheat, and potatoes from sugar beet; and images taken in September can be used to distinguish cereals from root crops and grassland, and potatoes from beet (Legg 1991). Therefore, the approach described here for rapeseed has wide application and could be adapted for quantifying gene flow from a variety of crop types in many countries.


This work was carried out under UK BBSRC/NERC Grant 45/G114202 (Brassica as a model to describe the pattern and consequences of transgene movement into the environment).