161 Kite Street (Locked Bag 21), Orange NSW 2800
2Charles Sturt University
Locked Bag 588, Wagga Wagga NSW 2678
PMB, Wagga Wagga NSW 2650
1Corresponding author – Phone: 02-63913195, Fax: 02-63913767
Reliable, up to date information on weed abundance, distribution, and change over time is essential for all aspects of broadscale weed management. Such information is necessary to evaluate control strategies, prevent spread to clean areas, and to improve weed management.
Conventional weed mapping techniques are expensive, time consuming and are generally not repeated frequently enough to monitor important changes in infestations. They are also inefficient where the target weeds cover a wide geographic area.
Remote sensing offers a low cost, repeatable alternative for mapping and monitoring weed infestations over large areas, although with several limitations. For remote sensing to be successful, the target weeds must have distinct reflectance differences from background vegetation, soil and stubble. For detection by current multispectral sensors, these differences must be great enough to compensate for the broad spectral bands and the pixel size of the sensor. Detection may also be limited by the density of the weed infestation.
Despite these limitations, mapping of infestations is possible. In a recent study, up to 80 – 86% of Scotch thistle infestations in pasture were detected across much of a Local Government Area, at an infestation density of down to 20% groundcover. Similarly, 72 – 82% of serrated tussock infestations were also detected, at an infestation level of down to 30 – 40% groundcover. However, in each case, attempts to map various weed density classes gave poor accuracy, due to confusion between the classes.
Such studies indicate the potential of remote sensing for weed mapping over large areas. The reliability of remote sensing for weed mapping will improve as imagery from new high resolution sensors becomes available.
The cost and importance of weeds to agricultural production means that accurate, up to date information on their location, density and spread over time is essential. Unfortunately, such information is rare in NSW (Campbell, 1991).
In part, this is due to the time and labour required for broadscale physical weed surveys. Data collection is slow, and difficulties in mapping exist where the weed population is large or has a broad geographic range. The variation in occurrence of annual weeds from year to year compounds these difficulties (Pitt and Miller, 1988; Auld, 1995). High cost prevents surveys being regularly repeated to determine change in infestations over time. Problems of accurate location of weed infestations may also occur where Global Positioning Systems (GPS) are not used (Lass and Callihan 1993).
Most information on weed distribution over NSW is at a coarse scale (eg. whole Local Government Areas), being collated primarily from questionnaires. Such data is inappropriate for management decisions, and only provides a very broad picture of infestations.
However, accurate maps of weed infestations and information on change at a farm, subcatchment and regional scale are extremely useful for all levels of weed management. Such maps are in high demand by Landcare groups, Weeds Officers, advisors and researchers.
Remote sensing has potential as an alternative means of weed mapping, being objective, cost efficient, regularly repeatable, and able to rapidly map large areas. Such regional scale analysis is extremely useful in monitoring weed spread and change over time, particularly the effectiveness of control strategies. Currently, remote sensing has however had only limited use for broadscale weed mapping, primarily due to the spectral and spatial limitations the current multispectral (many-band) satellite sensors and scene size limitations in using airborne sensors.
For multispectral remote sensing systems to be effective in mapping weeds:
- The target plant must have significantly different spectral reflectance (that is, the amount of light reflected to the sensor) from its background (stubble, soil or other vegetation);
- The width, number and position of the sensor’s spectral bands (spectral resolution) must be sufficient to detect these differences;
- The pixel size (spatial resolution) of the sensor must be sufficient to detect the weed infestation;
- The overpass cycle of the sensor must be sufficiently frequent to acquire data at key weed growth stages; and
- The data must be able to be processed and provided to end-users in a timely manner. (Fitzpatrick et al., 1990; Lamb, 1998; McGowen, 1998).
Remote sensing has often been used for mapping broad vegetation or land use classes, or general crop or pasture types. Mapping individual plant species is difficult, as plants rarely have ‘unique’, unchanging spectral reflectance characteristics (Fitzpatrick et al., 1990; Price 1994). Even where target weeds have spectral characteristics that differ greatly from other plants, changes can occur due to soil fertility, nutrient status, moisture status, disease or insect attack, and herbicide application. This complicates the use of remote sensing, particularly where it is used to map infestations over large areas (McGowen, 1998).
However, differences in plant size, growth rate, maturity, structure and colour at certain times of the year may aid discrimination. For example, major spectral differences often occur when target plants flower (Lass et al., 1996). An essential aspect of research into remote sensing of weeds is the identification of key stages of growth of weed species, to determine when they can best be discriminated from other vegetation. The growth characteristics of many weeds can assist in their identification by remote sensing. Woody weeds, due to their growth pattern, size and canopy characteristics, can often be readily discriminated (Everitt et al., 1992; Gardiner et al., 1998). The tendency of many pasture and crop weeds to grow in distinct patches, with few individual plants between, will aid their discrimination (Cardina et al., 1997; Rew et al, 1997).
Assuming remote sensing is appropriate for mapping a particular weed, the scale of mapping required and the spatial and spectral resolution of the sensor will determine the appropriate remote sensing system (either satellite or airborne).
Aerial photography has excellent spatial resolution, but has limited spectral resolution. Its high spatial resolution also carries a penalty in limited frame/scene size. Errors caused by distortion and vignetting (loss of brightness towards the edges of the frames) can also create difficulties (Jensen et al., 1986).
Most current multispectral (many-band) satellite sensors (eg. Landsat, SPOT, IRS) have relatively coarse spatial and spectral resolution. That is, they have a relatively large pixel size and a limited number of wide spectral bands, although they have the advantage of relatively large scene sizes and the imagery is relatively inexpensive. Consequently, they are most suited to sub-catchment to regional scale mapping. Airborne and newer satellite sensors (eg IKONOS) have better spatial and sometimes spectral resolution, but have the problems of limited scene size. They provide the capability to map at an individual paddock or property scale up to sub-catchment or catchment scale.
Despite these limitations, remote sensing has been successful at identifying and mapping a range of weeds. Aerial photography has been used in a number of studies, particularly for weeds with distinct, showy flowers (eg. capeweed – Arnold et al., 1985). Airborne multispectral imagery has been used to map a range of weeds in field crops (Brown et al., 1994; Lamb et al., 1999) and woody or herbaceous pasture and rangeland weeds (Everitt et al., 1995, 1996a, 1996b; Lass et al., 1996). Such imagery can be rapidly processed and used for precision spraying of paddocks, to reduce chemical application costs.
Fewer weeds have been successfully mapped at regional scales using broadscale satellite imagery. One notable example was in the mapping of moderate to dense infestations of bracken over Scotland, using Landsat Multispectral Scanner (MSS) imagery in a combined analysis with topographic and soils data (Miller et al, 1990). Most success at regional scale mapping has been obtained with woody weeds (Everitt et al., 1996a, 1996b; Gardiner et al., 1998).
The major limitation with current multispectral remote sensing systems is in discriminating light and scattered weed infestations. It is from such infestations that weeds most readily colonise new areas and extend their range, and these therefore represent the highest priority for control (Moody and Mack 1988). In most studies, the detection limit has been infestations of approximately 20 – 30% groundcover, even where airborne sensors are used with metre or sub-metre pixel size. However, this limit will vary with the particular weed and sensor. Problems are most likely where infestations have a similar size to the spatial resolution (pixel size) of the sensor (Lass et al. 1996; Lass and Callihan, 1997; Lamb, 1998; McGowen, 2000). Such limitations can be in part overcome by mapping areas of likely infestation or land cover types/associations at risk of infestation (Dewey et al., 1991; Peters et al.; 1992). Complex algorithms can also be used to assess whether each pixel is likely to contain a proportion of the target plant. The resulting risk maps can then be used as a guide to better direct field investigations.
Alone, analysis of remotely sensed data may never provide sufficient accuracy for the mapping of weeds. Ancillary information from traditional weed data collection, climatic, topographic, hydrographic and soils data many need to be incorporated with the remotely sensed data to provide the best results (Pitt and Miller, 1988; Fitzpatrick et al., 1990; Miller et al., 1990).
The future of high accuracy weed mapping will probably rely on the use of hyperspectral imagery. Hyperspectral sensors capture data over a very large number of extremely narrow spectral bands (often between 100 – 250). These sensors have limited availability at present, but offer the best potential for mapping light infestations of weeds, and in discriminating weeds with only subtle reflectance differences from other vegetation. The large number of spectral bands results in data sets of enormous size and complexity, which will present challenges for data handling and analysis. Also, scene sizes will be limited due to the high volume of data. However, such sensors will provide unique capability for identification of diverse vegetation types (McGowen, 2000).
The past success of remote sensing in mapping a range of crop, pasture and rangeland weeds despite its limitations shows the potential of the technique and ensures its greater use in the future.
A recent study investigated the potential and usefulness of Landsat Thematic Mapper (TM) imagery for mapping two major pasture weeds.
The target weeds
Serrated tussock (Nassella trichotoma) is the most serious pasture weed of NSW. It is extremely invasive, difficult to control and causes greater reductions in animal carrying capacity than any other pasture weed in Australia (Campbell and Vere, 1995). A broadscale survey in 1997 estimated light to heavy infestations as covering up to 900,000 ha of NSW, with scattered plants over a further 1.1 million ha (Jones and Vere, 1998). Scotch (Onopordum acanthium) and related thistles are major pasture weeds in the temperate high rainfall zone of south eastern Australia, and are rated as the second most significant thistle type in the NSW. Despite this, the extent of infestations is only known in broad terms, with the last (questionnaire) survey conducted in 1986 recording infestations of approximately 995,000 ha. The change in area over time and density of infestations is unknown (Briese, 1988; Briese et al., 1990).
Both weeds possess characteristics that aid their detection by remote sensing. For serrated tussock, the size of the tussocks, the perennial nature of the plants and the density of infestations assist discrimination. Large mature plants can be up to 60 cm high with a leaf spread of up to 75 cm, and infestation densities can reach 20 plants/m2. The distinct colour changes of the plants across the growing season are the major advantage for detection. The plants are brown through summer, and brownish green in autumn and spring. In winter, they are bleached yellow by frost. At flowering, the plants have a purple tinge, changing to a golden brown colour when the seedheads are elongating (Parsons and Cuthbertson, 1992; Campbell and Vere, 1995). These changes in colour are distinctly different from other species, particularly the genetically similar native spear and wire grasses.
Scotch and related thistles can act as annuals, biennials or short-term perennials, depending on the species and time of year of germination. They have distinct silvery-green leaves and whitish-grey stems (in part due to dense, white, woolly hairs present on the stems and leaves). Their distinctive colouration and large size of the plants (with rosette leaves of up to 40 cm long and stems of up to 2m in height) and the density of infestations (up to 5 flowering and many non-flowering plants/m2) favour their discrimination by remote sensing (Auld and Medd, 1987; Cavers et al., 1995; Parsons and Cuthbertson, 1992; Pettit et al., 1996).
Field spectral studies
Field spectral studies were carried out on both weeds and a range of other pasture and weedy plants across the 1997 growing season, to determine whether the reflectance characteristics of the plants (in the visible and near infrared areas of the spectrum) were sufficient to allow detection by remote sensing.
Serrated tussock was greatly different from most other vegetation when frosted in mid winter, and at flowering in early summer. However, in mid spring, its reflectance was similar to that of many native pasture species. Figures 1 and 2 show the spectral reflectance of the weed at frosting and flowering, in comparison to its reflectance in spring and to that of phalaris (which shows relatively ‘normal’ reflectance characteristics). Figure 1 also shows the position and width of the first four Landat TM bands (covering the blue, green, red and near infrared regions). The only difficulties in discrimination proved to be between the weed and frosted Kangaroo grass in winter (Figure 1), and hayed-off vegetation in winter and early summer (not shown), due to relatively similar reflectance characteristics in the visible area of the spectrum. Some discrimination was also possible between serrated tussock and other vegetation in mid spring, although this was much less reliable than at the other times of year.
Figure 1: Reflectance characteristics of serrated tussock when frosted, 1997
Figure 2: Reflectance characteristics of serrated tussock at flowering, 1997
The reflectance differences between Scotch thistle and other vegetation were also marked. The weed was consistently about 5 – 10% greater in reflectance over most wavelengths than most pasture and weed species across the 1997 growing season. Reflectance increased (particularly in the green and red areas of the spectrum) as the plant moved from the rosette stage to early bolting. Reflectance was dramatically higher at late flowering two years earlier (1995), measured in a preliminary study (Figure 3).
Analysis of the spectral data across the first four Landsat TM bandwidths, for each species and time of year, confirmed that both weed species could be discriminated at the key growth stages, but was best where data was combined across the growth stages. This demonstrated their potential for mapping using satellite imagery.
Figure 3: Reflectance of Scotch thistle across the growing season, 1997
Analysis of Landsat imagery for weed mapping
Extensive field data were collected over the 1997 growing season at sites in the Gunning and Boorowa Local Government Areas (LGAs), monitoring changes in ground cover of the target weeds at three infestation densities (light, moderate and heavy). A broader survey was used to collect information on the location of major land cover types (primarily native, naturalised, annual and perennial pastures, cereal crops and canola), as well as the location of infestations of the target weeds.
Landsat TM data was captured at three growth stages for each weed, matching the spectral study and field survey data. For serrated tussock, imagery was captured in mid winter (frosting), mid spring and at early/mid flowering. Imagery for Scotch thistle discrimination was captured on each of 3 consecutive Landsat overpasses in spring (16 days apart). These matched the late rosette, early bolting and early flowering stages.
For each species, the three dates of imagery were combined into a multi-date stack prior to analysis. The analysis of the spectral data and other studies indicated better discrimination was possible by such a method (Richardson et al., 1985; Lass et al., 1996). Analysis of the spectral data also suggested benefits from calculating ratios from certain bands, and adding these to the stack of imagery prior to analysis. This was done using the normalised difference vegetation index (NDVI).
The extensive field data were used to carry out a supervised classification of the imagery. In such a classification, the user locates known areas of various land cover types on the imagery, and uses these to ‘train’ the classifier to map the land cover on the rest of the imagery.
Results of the analysis
In both cases, the attempt to map three levels of weed infestation gave poor accuracy due to confusion between the classes. Discrimination of light infestations was poor, particularly for serrated tussock. However, aggregation of the infestation classes for each weed greatly improved the results, demonstrating that most of the confusion was between the three weed infestation classes themselves rather than between the weed classes and other land cover types.
Scotch thistle was most successfully mapped, with 80% of infestations being identified at a reliability of 97% (ie. only 3% of infestations were confused with other land cover types). The mapping of serrated tussock was less successful, but still proved reasonably good. Some 72% of infestations were identified, with a reliability of 87%. Unfortunately, very dry conditions occurred at the key frosting and flowering stages. The reflectance characteristics of (moisture stressed) serrated tussock at these stages were similar to those of other stressed or hayed-off vegetation, particularly in the visible area or the spectrum. Discrimination between serrated tussock and native grasses was a particular problem. Examples of the classification (and comparison with mapping by airborne sensors or aerial photography) are shown in Figures 4 and 5.
In both cases, the majority of error was in the identification and mapping of light infestations (less than 20-30% ground cover of Scotch thistle and less than 30-40% ground cover of serrated tussock). Removal of these classes from the analyses improved the accuracy of mapping, with 86% of the combined moderate-heavy Scotch thistle infestations identified, and 82% of moderate-heavy serrated tussock. The reliability of mapping of both species was similar to the previous analyses (although 5% less for serrated tussock).
Figure 4: Classification of Landsat imagery for serrated tussock infestations and comparison with approximate boundaries from aerial photography
Figure 5: Classification of Landsat imagery for Scotch thistle infestations, and comparison with visible infestations on airborne video imagery
Remote sensing has been successfully used for mapping weeds in a large number of studies covering a range of plant types and environments. The studies have shown that remote sensing can be successfully used at scales ranging from individual paddocks up to regional areas.
In this study, both target weeds showed distinct spectral differences during certain stages in their growth cycle that allowed differentiation from other vegetation. These qualities, coupled with high density of weed infestation, allowed mapping using Landsat TM imagery. Scotch thistle was more readily mapped (and at a lower infestation density) than serrated tussock, although light and scattered infestations of both weeds proved difficult to discriminate.
This has proven a general difficulty with weed mapping using multispectral imagery, as has the detection of species that do not show distinct spectral differences from other vegetation. These difficulties are likely to remain unresolved until imagery from hyperspectral sensors is readily available.
Some of these difficulties have been overcome by the combination of remotely sensed and ancillary data (such as soil, topographic and climatic information). This may still represent the best means to accurately map weeds, even when high spatial and spectral resolution imagery becomes readily available.
Despite the problems, and given the cost, time and labour requirements of conventional field survey, remote sensing represents a viable alternative to conventional mapping. With continual improvements in spectral and spatial resolution occurring, remote sensing is sure to play an increasing role in weed mapping.
Funding for this project was provided by the Rural Industries Research and Development Corporation, and is gratefully acknowledged. NSW Agriculture provided supplementary funding. Charles Sturt University and CSIRO Land and Water provided technical support and in-kind contributions, and further in-kind funding was provided by Upper Lachlan Advisory Services and Landcare. The assistance of the cooperating landholders and Weeds Officers in the Boorowa, Yass and Gunning LGAs is also gratefully acknowledged, as is the assistance of Helen Nicol (NSW Agriculture) for statistical support.
Arnold, G.W., Ozanne, P.G., Galbraith, K.A. and Dandridge, F. (1985). The capeweed content of pastures in south-west Western Australia. Australian Journal of Experimental Agriculture, 25 (1), 117-123.
Auld, B.A. (1995). A brief overview of weed survey methods. In Survey Workshop, Proceedings of a workshop, Wagga Wagga, NSW, December 1995, eds. D. Lemmerle and S. Corey. Cooperative Research Centre for Weed Management Systems: Adelaide, SA, Australia, pp 3-4.
Auld, B.A. and Medd, R.W. (1987). Weeds - an illustrated botanical guide to the weeds of Australia. Inkata Press: Melbourne, Australia.
Briese, D.T. (1988). Weed status of twelve thistle species in New South Wales. Plant Protection Quarterly, 3 (1), 133-141.
Briese, D.T., Lane, D., Hyde-Wyatt, B.H., Crocker, J. and Diver, R.G. (1990). Distribution of thistles of the genus Onopordum in Australia. Plant Protection Quarterly, 5 (1), 23-27.
Brown, R.B., Steckler, J.-P.G.A., Anderson, G.W. (1994). Remote sensing for identification of weeds in no-till corn. Transactions of the American Society of Agricultural Engineers, 37 (1), 297-302.
Campbell, M.H. (1991). Weed control in pastures - are we winning? Plant Protection Quarterly, 6 (2), 55-63.
Campbell, M.H. and Vere, D.T. (1995). Nassella trichotoma (Nees) Arech. In The Biology of Australian Weeds Vol 1. eds. R.H. Groves, R.C.H. Shepherd, R.G. Richardson and F.D. Panetta. R.G. & F.J. Richardson: Frankston, Victoria, Australia, pp.189-202.
Cardina, J., Johnson, G.A. and Sparrow, D.H. (1997). The nature and consequence of weed spatial distribution. Weed Science, 45 (3), 364-373.
Cavers, P.B., Groves, R.H. and Kaye, P.E. (1995). Seed population dynamics of Onopordum over 1 year in southern New South Wales. Journal of Applied Ecology, 32, 425-433.
Dewey, S.A., Price, K.P., and Ramsey, D. (1991). Satellite remote sensing to predict potential distribution of dyers woad (Isatis tinctoria). Weed Technology, 5 (3), 479-484.
Everitt, J.H., Escobar, D.E., Alaniz, M.A., Villarreal, R. and Davis, M.R. (1992). Distinguishing brush and weeds on rangelands using video remote sensing. Weed Technology, 6 (4), 913-921.
Everitt, J.H., Anderson, G.L., Escobar, D.E., Davis, M.R., Spenser, N.R. and Andrascik, R.J. (1995). Use of remote sensing for detecting and mapping Leafy Spurge (Euphorbia esula). Weed Technology, 9 (3), 599-609.
Everitt, J.H., Escobar, D.E. and Davis, M.R. (1996a). Integration of airborne videography, GPS and GIS technologies for mapping plant species and insect infestations. In Proceedings of the 2nd International Remote Sensing Conference and Exhibition, 24-27th June 1996, San Francisco, California, U.S.A., 1, pp. 239-258.
Everitt, J.H., Escobar, D.E., Alaniz, M.A., Davis, M.R. and Richerson, J.V. (1996b). Using spatial information technologies to map Chinese Tamarisk (Tamarix chinensis) infestations. Weed Science, 44 (1), 194-201.
Fitzpatrick, B.T., Hill, G.J.E., Kelly, G.D. (1990). Mapping and monitoring of weed infestations using satellite remote sensing data. In Proceedings 5th Australasian Remote Sensing Conference, Perth, Western Australia, 8th-12th October 1990, pp. 598-601.
Gardiner, D.B., Tupper, G.J. and Dudgeon, G.S. (1998). A quantitative appraisal of woody shrub encroachment in Western New South Wales. Rangeland Journal, 20 (1), 26-40.
Jensen, J.R., Hodgson, M.E., Christensen, E., Mackey, H.E., Tinney, L.R. and Shartiz, R. (1986). Remote sensing inland wetlands: a multispectral approach. Photogrammetric Engineering and Remote Sensing, 52 (1), 87-100.
Lamb, D.W. (1998). Opportunities for satellite and airborne remote sensing of weeds in Australian crops. In Precision Weed Management of Crops and Pastures, ed. R.W. Medd and J.E. Pratley. Proceedings of a workshop, 5th – 6th May 1998, Charles Sturt University, Wagga Wagga, NSW, Australia. Cooperative Research Centre for Weed Management Systems: Adelaide, SA, Australia. 154pp.
Lamb, D.W., Weedon, M.M. and Rew, L.J. (1999). Evaluating the accuracy of mapping weeds in seedling crops using airborne digital imaging: Avena spp. in seedling Triticale. Weed Research, 39 (6), 481-492.
Lass, L.W. and Callihan, R.H. (1993). GPS and GIS for weed surveys and management. Weed Technology, 7 (1), 249-254.
Lass, L.W., Carson, H.W. and Callihan, R.H. (1996). Detection of yellow starthistle (Centaurea solstitialis) and common St. Johnswort (Hypericum perforatum) with multispectral digital imagery. Weed Technology, 10 (3), 466-474.
Lass, L.W. and Callihan, R.H. (1997). The effect of phenological stage on detectability of yellow hawkweed (Hieracium pratense) and oxeye daisy (Chrysanthemum leucanthemum) with remote multispectral digital imagery. Weed Technology, 11 (2), 248-256.
McGowen, I.J. (1998). Remote sensing - background to the technology and opportunities for mapping of pasture weeds. In Precision Weed Management of Crops and Pastures, ed. R.W. Medd and J.E. Pratley. Proceedings of a workshop, 5th – 6th May 1998, Charles Sturt University, Wagga Wagga, NSW, Australia. Cooperative Research Centre for Weed Management Systems: Adelaide, SA, Australia. 154pp.
McGowen, I.J. (2000). Remote sensing for mapping serrated tussock and scotch thistle in pastures. In Proceedings of the 10th Australasian Remote Sensing and Photogrammetry Conference, 21st – 25th August, Adelaide, South Australia, Australia (CDROM, Paper No. 60).
Moody, M.E. and Mack, R.N. (1988). Controlling the spread of plant invasions: the importance of nascent foci. Journal of Applied Ecology, 25, 1009-1021.
Miller, D.R., Morrice, J.G. and Whitworth, P.L. (1990). Bracken distribution and spread in upland Scotland: An assessment using digital mapping techniques. Australian Institute of Agricultural Science Occasional Publication No. 40, July 1990. Proceedings of an International Conference - Bracken 89, University of Sydney, NSW, Australia, 18-21 July 1989, eds. J.A. Thompson and R.T. Smith, pp. 121-132.
Parsons, W.T. and Cuthbertson, E.G. (1992). Noxious Weeds of Australia. Inkata Press: Melbourne, Australia.
Peters, A.J., Reed, B.C., Eve, M.D., McDaniel, K.C. (1992). Remote sensing of brome snakeweed (Gutierrezia sarothrae) with NOAA-10 spectral image processing. Weed Technology, 6 (4), 1015-1020.
Pettit, W.J., Briese, D.T. and Walker, A. (1996). Aspects of thistle population dynamics with reference to Onopordum. In Thistle Management Workshop, eds. T.L.Woodburn, D.T. Briese, and S. Corey, Plant Protection Quarterly 11 (Supplement 2), 232-235.
Pitt, J.L. and Miller, I.L. (1988). A review of survey techniques for the detection of weeds with particular reference to Mimosa pigra L. in Australia and Thailand. Plant Protection Quarterly, 3 (4), 149-155.
Price, J.C. (1994). How unique are spectral signatures? Remote Sensing of Environment, 49 (3), 181-6.
Rew, L.J., Miller, P.C.H. and Paice, M.E.R. (1997). The importance of patch spraying resolution for sprayer control. Aspects of Applied Biology – Optimising Pesticide Applications, 48, 49-55.
Richardson, A.J., Menges, R.M. and Nixon, P.R. (1985). Distinguishing weed from crop plants using video remote sensing. Photogrammetric Engineering and Remote Sensing, 51 (11), 1785-1790.