Monash University - Faculty of Arts

Arts Faculty » Geography & Environmental Science » Research » Gis » Public » Dbulman »

The Application of Remote Sensing Imagery to Weed Mapping: A case study for Paterson's curse (Echium plantagineum L.)

Dave Bulman

Abstract:

Traditional weed surveying methods have been subject to the time and cost limitations of dedicated field personnel, and the (often reluctant) cooperation of land owners. These limitations have often prevented the delivery of extensive mapping programmes required for the effective management of an ever increasing number of invasive species. Ullah and others (1989a) demonstrated that using Landsat TM imagery, offered the promise of broad scale weed mapping, supplementing, at significant cost and time savings, the tedious work involved with field survey methods.

Research for this thesis was established to address the challenge that remote sensing might offer greater effectiveness in the fight against the spread of weeds in the light of an improved technological position from that of 1989 and a greatly improved understanding of weed ecology, the latter resulting (at least in Australia) from the establishment in 1994 of the Cooperative Research Centre for Australian Weed Management (CRCAWM).

The aim was to build on the techniques of the earlier work on Paterson's curse referred to above. However, since commencing this project in 1995 difficulties were encountered in detection of the flowering weed in recent Landsat imagery making mapping of its extent not possible by this means. In an endeavour to find an explanation for this, data relating to temporal and spatial climate variations, soils, topography, land management, weed control and other factors were collected and analysed to determine if environmental and cultural factors might be contributing to this decline in weed detection. But no clear explanation for the lack of Paterson's curse patterns in Landsat data could be determined from an examination of these data while field-surveys were still indicating quite heavy flowering intensities.

An analysis of air-borne hyperspectral CASI data with two-metre spatial resolution and a better ability to spectrally discriminate the floral components, however, restored the ability to map Paterson's curse flowering infestations and provided some ability to make limited estimations of flowering intensity. This data revealed that, because of the more rigorous regime of compulsory weed management in recent years, there is a greater degree of patchiness in the plant's distribution, that was not evident from examination of the Landsat imagery. This is now thought be the main reason for problems of detection using coarser Landsat imagery.

This result now offers considerable scope for the development of a better model for using current technology to map Paterson's curse and understands its distribution patterns. Such a model would require environmental and cultural data at a much greater level of temporal and spatial detail than is currently available, but would be ideally situated to make use of remotely sensed data from the next generation of higher spatial and spectral resolution space-borne (or current air-borne) instruments.

During the course of the satellite image analysis, some oblique aerial photography (OAP) for the study area became available. Using suitable software for ortho-rectificfication it was possible to make use of identifiable features in the images to ortho-rectify these and test their usefulness as a substitute for field-based ground-truth. This technique might prove to be useful when OAPs were available at or near the time of satellite overpass, and where access to important infestation areas may be restricted, or where an attractive cost benefit may obtain over ground surveys.

Introduction

General

Because extensive surveys of weed populations is both intensive and expensive, in Australia present knowledge about the distribution of many weeds is limited to those regions where their occurrence has already had significant impact on agricultural productivity. Essential to a greater knowledge of the distribution of weeds and their potential to spread is an understanding of the environmental factors that govern their dispersal and establishment. While laboratory and site based ecological and autoecological studies provide important understanding about plant behaviour and response in relation to local environment (see for example the studies of Roark, 1955, on Onion weed (Asphodelus fistulosis) and that of Piggin, 1976, on Paterson's curse), it is only through broader regional studies that examine population dynamics over the full range of environmental variables that any assessment of factors implicated in the distribution, spread and eventually control of specific weeds can be made. The application of remote sensing and GIS technologies to examine broader scale patterns of distribution and analysis in relation the regional patterns of environmental factors is one cost effective approach to addressing this question. The important and complementary contribution made by RS/GIS to ecology has been emphasised by Smith et al. (1990):

"The next challenge is to extend ecological observations from local areas to larger regions based on the properties measured by the images. If this can be done, the images can become a unique vehicle for exploring regional patterns of vegetation abundance and character, providing new ecological insights."

Already there has been successful application of these techniques to regional habitat mapping in the United States (eg Breininger et al. 1991 and Pereira and Itami 1991) and in Australia ( Hibberd et al. 1990). While the aim of these studies has differed from that of weed mapping, in general the principles are the same. To date there has been little serious attempt to model weed distribution using spatial data in this country and the potential for applying techniques of spatial data analysis to broader scale weed surveys and environmental modelling remains largely untested. There is considerable scope for the application of RS/GIS methodology to better understand weed distribution and spread taking into account the spatial distribution of physical environmental factors and cultural land use practices. For example:

• It can show the general distribution of a particular weed in areas that have not yet been mapped from field survey,
• It can show the potential patterns of weed spread,
• It can be used to identify critical areas where protection measures might be implemented to prevent further infestation,
• It can show the types of environments and the types of land management practice that are most susceptible to infestation,
• It can be used to monitor the effects of particular land use practices on weed infestations, and conversely
• It can be used to monitor the effectiveness of control measures under a variety of circumstances.

It is the aim of this project to investigate the potential of RS and GIS technology for mapping weed distribution, and to develop spatial models relating key factors to weed distribution, dispersion and control.

Choice of weed and area.

Paterson's curse (Echium plantagineum) has been a problem weed of pastures for many years, and there is an active and ongoing program of investigation into its ecology and appropriate control measures. It is also a key weed for investigation within the CRCAWM and is ideally suited to investigation using remote sensing because of the distinctive spectral characteristics of Paterson's curse during flowering. The effectiveness of using remote sensing for regional mapping of Paterson's curse was demonstrated by Ullah et al. in 1989 in an area near Lake Hume. The methodology developed for that project would form the basis for refinements to the techniques and for incorporation into a GIS.

Initial investigation would concentrate on the immediate area of the 1989 investigation and, if the techniques provide a useful basis for evaluation, could be extended into neighbouring regions on both sides of the border, contingent upon the availability of the suitable data.

Paterson's curse distribution

Originating, by first being introduced, in south eastern NSW, Paterson's curse has spread mainly, through stock movements, throughout southern regions of Australia and is now a weed of national significance. It is widespread in south-east NSW, Victoria, south eastern South Australia and south-western Western Australia (Figure 1) and extends to a lesser degree into south-eastern Queensland and western New South Wales. Its significance lies its ability to heavily infest pastures (Figure 2), thus reducing pasture values and has been recorded as injurious to some stock.

Aims

The original aims of this project were:

1. To examine the potential of RS/GIS techniques to contribute to regional weed ecological and biogeographical studies;
2. To use RS and GIS to achieve a more cost-effective method for mapping the extent of Paterson's curse distribution, and to gain a greater understanding of the factors governing its distribution and spread;
3. To provide a sound methodological basis using GIS to evaluate the effectiveness of weed control measures.

However, these aims were going to prove more difficult to achieve than at first thought because Landsat images acquired from the start of this project in 1996 did not show the existence of Paterson's curse as clearly as that of the 1988 image. This meant that research needed to be directed towards investigating possible explanations for this, given that dense infestations of Paterson's curse still ocurred on properties in the area.

The aims now became less concerned with the development of an "operational" methodology for mapping Paterson's curse than with attempts to explain why this weed was not being seen in recent Landsat imagery. Several lines of enquiry presented themselves, which now became the main focus of the research and included addressing the following questions:

1. Landsat 5 was ageing and had been in service past its expected replacement date: could this be the reasons for a lowering of sensor sensitivity making Paterson's curse less readily detectable?
2. Were there seasonal factors that could explain variablity in the flowering response of Paterson's curse?
3. Have cultural factors - land or weed management - lead to a change or decline in weed infestation or flowering response?
4. In view of these factors, what requirements and considerations would be needed for the development of a useful and practical application for modelling and mapping weed distribution (such as Paterson's curse) using remote sensing and GIS?

Methods

General

The general approach to this project was to examine and possibly refine the methods for extracting from remotely sensed imagery the extent and, possibly, some measure of infestation density of Paterson's curse over the study area in north-east Victoria. Research needed to be focussed on attempts to find explantions for this. Working hypotheses included the possible demise of sensors in the Landsat 5 thematic mapper given that it was overdue for replacement when Landsat 6 failed, the effects of variable local and regional climate, changes in land use or farming practices, or changes and developments in weed management strategies. Some efforts were also made to examine weed distribution in relation to landscape factors through the development of some GIS layers such as elevation, slope and aspect.

With the difficulties experienced using Landsat imagery, including data from the relatively new Landsat 7 ETM+ data, access to some air-borne CASI data became available making it possible to carry out some preliminary analysis of this higher spectral and spatial resolution data for any improvement in weed detection.

A more detailed description of the work done, in point form, is discussed below with the results and conclusions that were reached.

An outline of procedures, results and conclusions.

Some of the major procedural stages that were carried out for this project included:

1. Re-examination of the RS data used in the study of Ullah et al. (1989); re-evaluation of the methods used for this study and refine the image analysis and classification. In the 1988 image for the area, almost any method of examining the image allowed distinctive expression of infested areas. For example the bands 3, 2 &1 combination shows Paterson's curse as the darker purple areas in Figure 3. When using the false colour band combination of 4, 3 & 2 the pattern is less distinctive (Figure 4) because of the greater overwhelming response from the infra-red band at this time of peak growth for most pasture species, including Paterson's curse, but is still somewhat visible as an orangey colour. Paterson's curse is more dramatically expressed when using the PCA (Figure 5) and de-correlation stretches (Figure 6).
   However, with all of these visualisations, and with the classifications tried, there remains some confusion about the separation of lightly infested areas and forest margins. It is true that Paterson's curse encroaches into these areas, especially where some forested areas on grazing properties are used as un-improved pasture.
2. The spectral properties of Paterson's curse, and especially the differences between the floral components and the vegetative parts (Figure 7) gave rise to the possibility of devising a band ratio method of distinguishing between Paterson's curse and other pasture species. As can be seen from the spectra, there is a lowering of reflectance in the green band for the floral copmponents. This was used in a ratio (Figure 8a): Where PCIndex refers to a "Paterson's Curse Index", and the Red, Green and Blue are the DN values for each of bands 3, 2 and 1 respectively, that allowed for a simple but effective single-class definition, with some improvement in discrimination between Paterson's curse and forest margins. Further improvements were achieved by using an empirically determined near-infrared threshhold value to separate some forested areas from the more vigorously growing pasture (Figure 8b) : Where x is an empirically determined near infra-red (NIR) band (band 4) DN value for the difference between the pasture and other forms of land cover. A NULL value is recorded in the output raster (PCindex) for NIR values less than or equal to x.
3. The present research commenced with acquisition of more recent satellite imagery and analyses using similar methods as those used for the 1988 image. However, none of the images from Landsat 5 (1996, 1997, 1998 & 1999) or from the more recent Landsat 7 (ETM+) 1999 imagery revealed as clearly the presence of Paterson's curse. Some indication of its presence in the more heavily infested areas could be discerned in the 1996 image but, because of heavy cloud cover, analysis and classification was not possible. Infested areas could not seen or analysed in any of the later images either using any of the above methods. This raised questions as to why analyses of these later images were proving unsuccessful.
4. Seasonal variations in some climatic factors were investigated. These included local rainfall records (temperature data for the area was too limited for any useful examination of this variable), Southern Oscillation Index (SOI) patterns were examined, and the Drought Index component of Fire Danger Index (Keetch and Byram 1968, Crane 1982) were calculated and examined for any significant patterns. Apart from some general trends that might be equated with anecdotal estimates of better or worse years for Paterson's curse, there were no definitive patterns found that might explain the demise in expression of floral spectral response in satellite imagery.
5. Next, there arose some concerns about sensor sensitivity of the Landsat 5 spectrometer. This sensor had been in continuous operation for about 14 years - much longer than would have been the ccase had Landsat 6 been successfully launched. Little official information about the status of the satellite or its sensors was able to be obtained. An experimental process was tried to spectrally match the band sensor reponses to that of the original 1988 image using a method of regression analyses, originally devised by Furby et al(1996) as a rigorous means of matching spatially separate images, using spectrally invariant land-surface features. While this uncovered some minor reduction in sensor sensitivity for the blue band primarily, the corrections did not assist in the detection of Paterson's curse.
   As a further test for issues that might be related to the Landsat 5 sensor, at the time Landsat 7 was launched, it was possible to obtain imagery from both satellites over identical areas but one day apart. This later image, in spite of its better spatial resolution and greater sensitivity, proved to be no better at indicating the locations of flowering Paterson's curse than the Landsat 5 image, even after subjecting it to the regression matching technique. On these bases, sensor issues were largely dismissed as a significant problem.
6. The acquistion of some high spatial resolution CASI data in 1998 provided a much better basis for examining the issues of Paterson's curse detection. This data in 14 bands (Table 1) was collected at 1 m resolution (instead of the intended 2 m resolution) and consequently did not cover some of the survey sites that were included in the study. Nonetheless, it covered enough infested areas to allow for an examination of the bands for possible use in an equivalent ratio method similar to that of the 1988 Landsat image. Immediately noticeable was the clarity with which Paterson's curse flowers showed up in a number of band combinations (see Figure 9 for example). This could be attributed to there being a greater spectral separation between the red and green bands in the overlap region between vegetative and floral components (Figure 7), and the clearer representation of flowers from what now seemed to be greater patchiness when viewed with the smaller Instantaneous Field of View (IFOV) of the CASI data.
7. Several combinations of bands were tried in a ratio configurations where ratios using bands 2 or 3 with 7 & 11 provided reasonable representations of infested areas (Figures 10a & b), however band 9 alone (673.5 - 686.1 nm) provided the means for obtaining an approximate correlation with Paterson's curse estimated flowering density (R² = 0.59), where it was above an estimated 2%(Figure 11), along a measured transect in some infested properties. Properties with lower infestation levels or having a greater degree of patchiness showed a much lower correlation.
8. Conventional classification methods (eg ISOClass & Maximum Liklihood) proved to be as elusive in defining Paterson's curse in this imagery as had been the case for Landsat images. However this imagery lent itself to a classification method known as Spectral Angle Mapper (SAM) in which a target class (or classes) are defined by a vector in multi-dimensional spectral space (see for example McCloy 1977) which has an implementation in the ENVI image analysis package as part of the hyperspectral analysis routines (Research Systems Inc. 1988). The results produced by using SAM provided a much clearer picture of the infested areas (Figure 12 - The lettering indicates farm property boundaries) and a possible explanation for the difficulties experienced when using Landsat data.
9. As can be seen in Figure 12, there is much greater degree of patchiness in the patterning of infestations at the "paddock" scale than is evident from the Landsat imagery. While this observation doesn't, of itself, provide a causal explanation, it can be speculated that, and anecdotal evidence seems to support, the notion that the increased need to control weeds under recent legislative imperatives has meant that Paterson's curse is having less opportunity to grow into denser stands.
10. The explanation for lack of detectability, it would seem, arises from the observation that in the coarser (30 m) resolution Landsat imagery there is a much greater dilution of the floral signal (Figure 13a) by the background pasture vegetation. At higher spatial resolutions achieved through the use of CASI data (Figure 13b), at least some pixels will consist of a majority of flowers and therefore consitiute the greatest reflectance.
11. Part of the difficulties in determining the outcomes from an image analysis is the time lag that occurs between analysis and field validation. Often, and particularly in actively grazed pastures, this time lag can mean quite considerable changes can take place in plant and field conditions between the time of acquisition of imagery and when field verification can be undertaken. To some extent these temporal changes can be anticipated by knowing areas that are likely to be useful for ground-truthing and field work can be carried out at or around the time of image acquisition. This was possible, for example, when planning the CASI data mission for the transect data to be collected .
12. In other circumstances this may be more difficult. As a substitute for field ground-truthing, an example has been given in this thesis where some obliques aerial photographs were taken from an aircraft on one occasion at about the same time as image acquisition was being anticitated. While the example shown in Figure14 may not provide the ultimate in accuracy of mapping the existence of Paterson's curse, nor in evaluating its extent, it does demonstrate a potential for testing the validity of remote sensing image analysis where other methods may not be possible. With forward planning, the acquisition of aerial photographs may provide a useful adjunct to convetional analytical methods and also provide some improvement in quatitative measures of infestation level. There is scope for further work in this regard.
13. Of greatest significance in outcome from this work is the difficulties that need to be addressed when applying remote sensing analysis to issues, such as weed mapping, that embrace multiple areas of research and knowledge to understand the problem. In this work there has been the realisation that three main areas of research are involved. These, in broad terms are:
    1. The biological and ecological properties of Paterson's curse,
    2. An understanding of the characteristics of remote sensing, sensors and platforms, and
    3. The physical and cultural setting in which these operate,
    and the interactions between these that will influence the ability, reliability and effectiveness of a particular methodology in mapping an predicting the occourrence of a particular weed.
14. In order to develop such a model for remote sensing, there would be a need to obtain data relating to the physical environment (eg soils, topography, climate, natural vegetation) and the cultural environment (eg property boundaries, landuse, land management, history of Paterson's curse infestation, disturbance factors) to build a GIS database. This data may need to be quite refined in temporal and spatial contexts. The model would also need to address issues related to remote sensing with regard to the properties of the target species. For example the nature of the spectral responses, the influences of spatial, spectral and radiometric characteristics, the platforms available and the relative cost for image acquistion, to mention just a few. And, finally, a thorough understanding is needed of the target species, its phenology, propogation mechanisms, spectral properties that make it distinctive, how these change over time as well as how plant responses to various environmental (eg soils, climate) and cultural (eg grazing, spraying control and spread) factors can affect distribution patterns. A possible configuration for such a model may look like Figure 15.

References

• Breininger, D. R., Provancha, M. J. & Smith, R. B., 1991. Mapping Florida Scrub Jay habitat for purposes of land-use management,Photogrammetric Engineering and Remote Sensing, 57(11):1467-1474.
• Busby, J. R., 1991. BIOCLIM - a bioclimatic analysis and prediction system, In: Margules, C. R. & Austin, M. P. (eds), Nature Conservation: Cost Effective Biological Surveys and Data Analysis , CSIRO, Melbourne. pp 64-68.
• Crane W J B, 1982. Computing grassland and forest fire behaviour, relative humidity and drought index by pocket calculator. Australian Forestry. 45(2):89-97.
• Davis, F. W., Quattrochi, D. A., Ridd, M. K., Lam, N. SÐN., Walsh, S. J., Michaelson, J. C., Franklin, J., Stow, D. A., Johannsen, C. J. & Johnson, C. A., 1991. Environmental analysis using integrated GIS and remotely sensed data: Some research needs and priorities, Photogrammetric Engineering and Remote Sensing, 57(6):689-697.
• Hibberd, J. K., O'Neill, A. L., Marthick, J. & Sim, R. L., 1990. Integration of satellite imagery and geographic information systems for habitat mapping and the prediction of fauna distributions: preliminary results and future directions, Proceedings of The Ecolological Society of Australia, 16:449-457. Pereira, J. M. C. & Itami, R. M., 1991.
• Furby, S. L., Palmer, M. J. & Campbell, N. A., 1996. Image calibration to like values. Proceedings of the 8th Australian Remote Sensing Conference . Canberra (Electronic edition).
• Keetch, J. J. & Byram, G. M., 1968. A drought index for forest fire control. U.S.D.A Forest Service Research Paper SE-38.
• Pereira J M C & Itami R M, 1991. GIS-based habitat modeling using logistic multiple regression: A study of the Mt. Graham Red Squirrel,Photogrammetric Engineering and Remote Sensing , 57(11):1475-1486.
• McCloy, K. R., 1977. The Vector Classifier. Proceedings of the Eleventh International Symposium on Remote Sensing of Environment. Environmental Research Institute of Michigan, Ann Arbor. Vol 1 pp535-543.
• Morley, T. and Stapleton, P., 1999. The Patersonís Curse Management Handbook. Department of Natural Resources and Environment, Victoria. 41pp.
• Piggin C M, 1977. The hebaceous species of Echium (BOROGINACEAE) naturalized in Australia. Muelleria. 3(4):215-244.
• Piggin, C. M., 1976. The Ecology and Control of Paterson's Curse (Echium plantagineum L .), Unpublished PhD Thesis, University of Melbourne.
• Piggin, C. M. & Sheppard, A. W., 1995. Echium plantagineum L., In: Groves, R. H. & Shepherd, R. C. H. (eds), The Biology of Australian Weeds Vol 1, R. G. & F. J. Richarson, Melbourne. pp87-110.
• Research Systems Inc., 1998. ENVI Tutorials. Tutorial.pdf.
  URL : http://www.rsinc.com/envi/tutorial.pdf
• Roark, B., 1955. The Autoecology of Asphodelus fistulosis, Unpublished PhD thesis, University of Adelaide.
• Smith, M. O., Ustin, S. L., Adams, J. B. & Gillespie, A. R., 1990. Vegetation in deserts: i regional measure of abundance from multispectral images, Remote Sensing and Environment, 31:1-26.
• Ullah, E., Shepherd, R. C. H., Baxter, J. T. & Peterson, J. A., 1989. Mapping flowering Paterson's Curse (Echium plantagineum ) around Lake Hume, north eastern Victoria, using Landsat TM data, Plant Protection Quarterly, 4(4):155 - 157.