Transcript Document
Using Airborne Remote Sensing in GM Risk Assessment Luisa Elliott, Dave Mason Joel Allainguillaume & Mike Wilkinson AIM: To model gene flow from oilseed rape to Brassica rapa (a wild relative) on a national scale What is GENE FLOW? Gene flow is the movement of genes between populations CROP CROP WILD RELATIVE Why does gene flow need to be modelled? •The safety of Genetically Modified (GM) crops is the subject of much debate worldwide. •For a hazard to occur from the movement of transgenes a series of events must occur. •Gene flow represents the first step and can be measured by the formation of hybrid plants. Current gene flow model •Landsat images used to detect the oilseed rape •B. rapa grows mainly on riverbanks – historical information was combined with ground survey data to work out which river systems contain the wild relative •Oilseed rape fields that grew next to waterways were identified and a spatial model of gene flow was developed Hybrid numbers •Amount of oilseed rape growing next to waterways was combined with river survey data to enable estimation of hybrid numbers •Number of hybrids = 26,000 per annum in the UK •BUT … this prediction has a large confidence range of 22,000 … •The large error margin is mainly due to uncertainties in the distribution of B. rapa •138,000km of banks predicted to contain B. rapa •We have surveyed about 500km of river and canal banks by boat and foot (less than 1% of total!) •The tributaries are generally not accessible •Therefore, we need to assume that the surveyed areas represent all waterway banks Airborne remote sensing provides a useful tool for surveying inaccessible places and at a much faster rate than field work NERC-ARSF obtained approx 85km of both ATM and CASI –2 data in May 2003 (7 flight lines). Hypotheses to test 1. B. rapa occurs on the tributaries in the same frequency as along the main rivers 2. B. rapa seeds are carried in waterways and dispersed onto the banks during flooding events How to test the hypotheses •Compare the distribution of B. rapa along the banks of main rivers with that of their tributaries •Compare the distribution of B. rapa along river banks with that of along canal banks (canals do not flood) First step is to identify B. rapa in the aircraft images •Unsupervised K-means classification (in conjunction with ground survey information) successfully used to detect the larger populations (>2m x 2m) •Populations detected more accurately in the ATM images than the CASI i.e. extra spectral information more important that increased spatial resolution •Unsupervised K-means classification is not suitable for all images because it is not possible to obtain sufficient ground reference data for all flight paths. •Need to calibrate images, to enable classification of one image using ground reference data from a different image. •Cross-track illumination correction function in ENVI used to correct for reflectance differences across the width of each swath •Flat field correction then carried out to correct for differences between images (dark water pixels used to represent the flat field pixels). An example of an image before and after cross-track and flat field calibration •After calibration, a supervised maximum likelihood classification was used for one image using spectral information from a separate image to train the classifier. •This method was tested for an image in which all B. rapa positions were known and resulted in the detection of 94% of the correct total number of plants. Therefore, large B. rapa populations can successfully be identified in the ATM images, even if we have no ground reference data for that image. But what about the smaller populations? The large populations account for 95% of the total plant numbers and 14% of the total population numbers Matched filtering (in ENVI) can be used to detect smaller population (>1m x 1m) •Combining results of matched filtering with results of large population classification enables detection of 99% of the plants and 79% of the populations (of at least 1m x 1m in size) •Populations >1m x 1m for 98% of plant numbers and 32% of population numbers. Progress so far … •Cross-track illumination correction followed by flat-field calibration can be used to normalize images and enable classification of one image using spectral information from an independent image •Supervised maximum likelihood classification can detect large populations (>2m x 2m) •Matched filtering can be used to detect smaller populations (>1m x 1m) Still to do … •Classify all images and look at overall distribution of B. rapa. •Test the hypotheses and update current model of gene flow.