Environmental Monitoring and Assessment (2006) 117: 505–518 DOI: 10.1007/s10661-006-0768-3
c Springer 2006
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION IN SENSOR BASED PRECISION AGRICULTURE ¨ HANS-R. LANGNER∗ , HARTMUT BOTTGER and HELMUT SCHMIDT Institute of Agricultural Engineering Bornim (ATB), Dept. Eng. for Crop Production, Max-Eyth-Allee 100, 14469 Potsdam, Germany (∗ author for correspondence, e-mail:
[email protected])
(Received 23 December 2004; accepted 5 July 2005)
Abstract. Many technologies in precision agriculture (PA) require image analysis and image- processing with weed and background differentiations. The detection of weeds on mulched cropland is one important image-processing task for sensor based precision herbicide applications. The article introduces a special vegetation index, the Difference Index with Red Threshold (DIRT), for the weed detection on mulched croplands. Experimental investigations in weed detection on mulched areas point out that the DIRT performs better than the Normalized Difference Vegetation Index (NDVI). The result of the evaluation with four different decision criteria indicate, that the new DIRT gives the highest reliability in weed/background differentiation on mulched areas. While using the same spectral bands (infrared and red) as the NDVI, the new DIRT is more suitable for weed detection than the other vegetation indices and requires only a small amount of additional calculation power. The new vegetation index DIRT was tested on mulched areas during automatic ratings with a special weed camera system. The test results compare the new DIRT and three other decision criteria: the difference between infrared and red intensity (Diff), the soil-adjusted quotient between infrared and red intensity (Quotient) and the NDVI. The decision criteria were compared with the definition of a worse case decision quality parameter Q, suitable for mulched croplands. Although this new index DIRT needs further testing, the index seems to be a good decision criterion for the weed detection on mulched areas and should also be useful for other image processing applications in precision agriculture. The weed detection hardware and the PC program for the weed image processing were developed with funds from the German Federal Ministry of Education and Research (BMBF). Keywords: decision criterion, image processing, mulched cropland, red threshold, signum function, spectral sensing, vegetation index, weed detection Abbreviations: DIRT: difference index with red threshold; LAI: leaf area index; NDVI: normalized difference vegetation index; NIR: near infrared; WDRVI: wide dynamic range vegetation index
1. Introduction Many production steps in sensor based precision agriculture (PA) require a differentiation between weed and soil. Capturing image data by cameras and the appropriate data processing are technical solutions for the detection of weeds, including the decision between weed and soil. The automatic differentiation between weed and soil can be achieved by the use of a vegetation index as the decision criterion (Baret and Guyot, 1991). The changes of the soil and the optical soil properties can be adjusted
506
H.-R. LANGNER ET AL.
using a soil adjusted vegetation index (Rondeaux et al., 1996). The differentiation process between weeds and soil uses the differences in the spectral properties of weeds and soil. The reflectance value of green vegetation is very low in the red optical range around 670 nm. Weeds exhibit the highest spectral absorption in the red region, but strongly reflect incident irradiation in the near infrared range (NIR). Soil has only a small difference between reflectance in the red and in the near infrared range. The distinct contrast in the spectral behaviour of weeds and soil is well known. Many techniques have been attempted in the literature to detect weed seedlings and to monitore the temporal and spatial variations of weed densities (Biller and Schicke, 2000; Chapron et al., 1999; Lamb and Brown, 2001; Perez et al., 2000). Capturing of weed images and the analysis of weed scenes in a computer are prerequisites for precise herbicide applications or for non-chemical weeding technologies with sensor control. The image processing technology in general and the accuracy of the weed detection determine the success of weeding. The sensor approach can be used to adapt the amount of herbicides to the spatial amount of weeds (Wartenberg and Dammer, 2001; Marchant et al., 2001). The application of herbicides has potentially negative environmental impacts depending on the application technology. New technological developments are necessary with the goal of herbicide treatment optimisation. One part of the technological improvements are the development of new camera sensors and the introduction of new image processing software. The sensor approach includes two steps: Detecting and rating the weeds on croplands and using the weed rating results for direct control of the herbicide treatment (Aitkenhead et al., 2003; Tian, 2002; Christensen and Heisel, 1998; Felton and McCloy, 1992). The mulching of cropland is often used as a production step for slowing down the development of weeds. On mulched croplands the soil is mixed with plant residues and straw. From the image processing point of view the mix of plant residues, straw and soil is called background. The image-processing task on mulched cropland is complicated because the properties of all materials influence the decision between weed seedlings on one site and the background (plant residues, straw and soil) on the other site. The conventional image processing technology needs to be adapted to the sensor approach for mulched areas. The sensor approach for mulched cropland should use an appropriate decision criterion for the differentiation between weed seedlings and background. The main task during a weed image analysis is the decision between active objects (weed) and background. The background gives no information of interest, so it can be discriminated. To achieve this, the software has to evaluate a decision criterion for each pixel and to separate weed pixels and background pixels. The vegetation index, based on red and infrared reflectance values, performs as the decision criterion for each pixel. Furthermore the number of weed pixels summarized for each weed scene can be used as the output variable of the weed image processing. In general the weed leaf area index of each weed
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
507
scene can be used as a control parameter in PA weeding technologies (Wartenberg and Dammer, 2001). The quality of the weeding process is strongly determined by the weed detection technology and especially by the selectivity of the type of vegetation index used as decision criterion. Fotometric cameras used for weed rating applications combine one optical channel in the red spectral region and a second optical channel in the infrared region (Gerhards et al., 2002). Many techniques have been attempted in the literature to detect weed seedlings on soils (Chapron et al., 1999; Biller and Schicke, 2000; Lamb and Brown, 2001; Marchant et al., 2001; Perez et al., 2000). The differentiation between weed pixel and background pixel uses the difference in spectral properties between vegetation and the background material. The reflectance values of green vegetation are very low in the red range around 670 nm. So weeds exhibit the highest spectral absorption in the red region. Their reflectance intensities are usually below 10% in the red region. In the near infrared range (NIR) at around 780 nm wavelength, green vegetation strongly reflects incident irradiation. The photo-synthetically active area ends in the infrared region, so the absorbance is low and the reflectance intensities NIR exhibit high values of about 50% in this region. This distinct contrast in the spectral behaviour of vegetation is widely used (Buschmann and Nagel, 1993). A number of spectral vegetation indices have already been devised for detecting and rating of weeds or crops. One classification method to get a differentiation criterion between weed and background is based on the ratio of the two intensity values IR and R. The quotient IR/R varies strongly and reaches values of 20 or more for healthy weeds. The IR/R ratio is often shifted to lower values, called soil adjustment of the vegetation index. Normally the red intensity values R are below 0.1, but for soil compensation the values are shifted to the range 0.15 to 0.3 (Oppelt and Mauser, 2004). The following equation is used for the comparison with other vegetation indices: Quotient = NIR/(R + 0.15)
(1)
The adjustment of the quotient IR/R in (1) and numerical problems with high IR/R ratios can be avoided by using the simple difference between the IR and R values as the vegetation index called Diff. Diff = NIR − R
(2a)
By using the difference Diff, normalized to the sum of NIR and R, the formula is the well-known Normalized Difference Vegetation Index NDVI (Baret and Guyot, 1991). The NDVI difference between NIR and R is very often used as the decision criterion. NDVI = (NIR − R)/(NIR + R)
(2)
508
H.-R. LANGNER ET AL.
Despite its extensive use the NDVI has some disadvantages, namely the non-linear relationship between the parameters NDVI and leave area index LAI (Buschmann and Nagel, 1993) and the poor correlation results between NDVI and the aboveground biomass (Baret and Guyot, 1991). The relationships between the NDVI and other parameters like the LAI or the biomass have important consequences for the weed ratings and for the monitoring of weed dynamics on tilled areas. To minimize the non-linearity between NDVI and LAI, the Equation (2) can be advanced with a weighting coefficient a, with values for a in the range 0.1 to 0.2 (Gitelson, 2004). The resulting equation for a Wide Dynamic Range Vegetation Index WDRVI (Gitelson, 2004) can be used for better correlation results: WDRVI = (a ∗ NIR − R)/(a ∗ NIR + R)
(3)
The WDRVI uses the weighting coefficient for stretching the NDVI curves and to avoid saturation effects between WDRVI and the biomass. The WDRVI is a new vegetation index (Gitelson, 2004) and needs further testing in combination with LAI and biomass measurements. The WDRVI is not included in the weed rating evaluations of this article, because we have no results using the parameter a for stretching the NDVI curves. All vegetation indices compare information of two or more spectral bands. In our investigation we used the red spectral band R and the near infrared band NIR. The goal of our investigation is to find a vegetation index that gives the highest reliability for the weed/background differentiation on mulched cropland. The steps to reach the goal were spectral reflectance measurements of materials from mulched croplands in the laboratory and during tractor based weed ratings. The evaluation of the experimental results includes the comparison between three vegetation indices and the new criterion, the difference index with red threshold (DIRT). The article compares the quality of the vegetation indices with the help of a quality parameter for weed detection on mulched cropland. In the discussion we argue, that the vegetation index DIRT is important for the accuracy of weed ratings as well as for productivity improvements in PA practice.
2. Methods First of all the spectral reflectance of some material samples collected from a mulched cropland were measured in the ATB laboratory. The reflectance spectra’s of the material samples (straw, weed and foliage) from a mulched area are shown in Figure 1. The reflectance curves of weed seedlings have a typical behaviour with a sharp minimum in the red region. The reflectance curves of straw parts are quiet different and show medium to high values in the red region. The detection of weeds on
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
509
Figure 1. Spectral reflectance of arbitrary materials from mulched plants.
mulched areas is difficult by using only the NDVI as the decision criterion, because of the unspecific reflectance curves of straw material. The difference IR − R can yield high values for straw material, delivering false decisions (straw parts as weed seedlings). The detection of weed seedlings on mulched areas is a difficult task, if the detection rule is only based on the difference between infrared and red intensities. The weed detection process on mulched cropland needs an advanced vegetation index that includes the medium to high red intensity values of straw in the decision criterion. Evaluations of the laboratory measurements and starting from the basic definition of the NDVI, a vegetation index with extended parameters has to be derived for weed ratings on mulched cropland. The detection of weed seedlings between straw materials is a difficult task, especially when the weed/background decision is based only on NDVI calculations. Some parts of straw on mulched areas have high red and very high infrared reflectance values. Other kinds of straw have only medium red reflectance values, but the difference IR − R is still high. The decision rule for weed has to take the red reflectance value R into account, yielding the equation for a new vegetation index. The NDVI is focused on the difference between infrared and red reflectance. The raw reflectance value R is not directly included in the NDVI equation. The formula for a new vegetation index with extended parameters includes the threshold value β for the red reflectance R in such a way, that the new classification can distinguish between materials with medium/high red reflectance and materials with low red
510
H.-R. LANGNER ET AL.
Figure 2. The NDVI as a function of the red intensity R.
reflectance values. The new vegetation index that gives the higher detection reliability for weeds on mulched plants points out, that the measured intensity value R has to be below the threshold value β. For weed detection we choose a red threshold value β that is above the maximal red reflectance of the weed species. If the measured intensity R of an arbitrary image pixel is higher then the threshold β, the sign of the NDVI value is inverted (switched to the opposite sign), indicating that this pixel is not a weed pixel. With the signum function sign( β − R) it is possible to combine the new decision criteria and the NDVI in one formula, called the Difference Index with Red Threshold (DIRT): DIRT = sign(β − R) × (NIR − R)/(NIR + R)
(4)
Plots of the NDVI as a function of the red value R are shown in Figure 2 and plots of the new DIRT as a function of the red reflectance R are shown in Figure 3. The sharp cut of the plots in Figure 3 stands out due to the threshold value of 0.12, a value that was used for the threshold β in weed rating tests. During our test the value β was fixed to 0.12 and never changed. But under special circumstances, depending on the materials of the mulched areas, this threshold may be adjusted in the range from 0.08 to 0.15.
3. Realization of the Measurements The hardware for the weed rating tests consists of a special camera, a frame grabber card and the embedded version of a personal computer. The camera is a MS2100 CIR (by distributor Laser2000, Germany) with spectral filtering for red and infrared
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
511
Figure 3. The DIRT as a function of the red intensity R.
radiation. The camera delivers two spectral images of a weed scene to the frame grabber at one time (Laser2000, 2003). During the tests the camera was mounted in front of a tractor in a height of about 40 cm, so the weed scene is 20 cm × 15 cm when using the whole image area. The camera was oriented with the longer image side in relation to the drive direction. Prior to weed rating, the user can choose the length and the frequency of the image series (that is like a gap free digital film). If the user chooses a frequency of 15 images per seconds with 5 images per meter the maximum weed rating speed can be 3 m/s. The example of a weed scene is shown in Figure 4. The left part is the original image of the weed scene and the right part of Figure 4 is the corresponding binary image after background discrimination using the DIRT index.
Figure 4. Weed-rating scene (left) and corresponding binary image with suppressed background (right).
512
H.-R. LANGNER ET AL.
The software program “Beikraut” controls the entire data flow out of the weed camera MS2100, through the Matrox frame grabber card and into the data processing software in the embedded PC. The software “Beikraut” is programmed as a modular solution. The software module Matrox Imaging Library 6.1 controls the frame grabber hardware Meteor2, the driver and the dynamic link-libraries of the manufacturer Measurement Computing Corp. uses the DAS-6025 measuring card. The software modules are object oriented designed, using the programming language C++, and interact on a message transfer basis. After capturing and transferring the images into the personal computer, the software routines compare the intensity values for each pixel of the red and infrared image and calculate the decision criterion for the weed detection. The summary of the calculation are the weed count value and the weed leaf area index LAI, which can be used for advanced data processing or as a direct control value for the herbicide application technique. A sensor wheel is installed at the tractor, which delivers a path signal for the image processing software. The path signal can be entered manually because of the fact that the same weed rating has often to be repeated several times. The user can enter the true path value and the calibration software uses the value for partitioning the path and for assigning the number of images. To test the new DIRT decision criteria, 42 material samples were collected from a mulched plant. The first group of test materials includes 21 pieces of 3 kinds of weed (7 pieces of ripwort-plantain/upper site, 7 pieces of ripwort-plantain/lower site and 7 pieces of dandelion). The second group of test materials includes 21 pieces of straw (7 pieces of pea straw, 7 pieces of wheat straw/upper site and 7 pieces of wheat straw/lower site). The material samples were spatially distributed in 7 groups, were each group contains the following fixed order of the six test materials: ripwort-plantain (upper site), ripwort-plantain (lower site), dandelion (upper site) and wheat straw (lower site), wheat straw (upper site), pea straw. Seven groups of test materials were place on a mulched test area in a distance of one meter and a number of weed images were captured during the test drive. The results are time series of red intensity images and of infrared intensity images simultaneously recorded during the test-drive. The test software “Beikraut” stores the red and infrared images and selects a pair of red and infrared values for each of the 42 material samples. The red and infrared intensity values of all 42 samples in the 7 spatially distributed groups were selected from the images, captured during the weed-rating. For the speed of the image processing with high pixel resolution, it is important that the vegetation indices of Equations (1), (2) and (4) take only a small amount of computational power for the calculation.
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
513
Table I summarizes all 42 red intensity values and all 42 infrared intensity values as a function of the six test materials.
4. Results The raw intensity values (IR, R) of the weed and straw samples in Table I were measured during an automatic weed-rating test. The test software “Beikraut” calculates various kinds of decision criteria for each image pixel. For the weed-rating test we choose the decision criterions: – – – –
the simple difference between red and infrared intensities, the soil adjusted quotient (Equation (1)), the NDVI (Equation (2)), the new DIRT (Equation (3)).
Figures 5–8 show all the calculated vegetation indices based on the intensity values of the test. For each test material the median of the seven vegetation- index values is plotted as bar graph. To summarise the results, the definition of a quality parameter Q is necessary for comparing the different kinds of vegetation indices. For checking the decision quality between weeds and straw we choose the minimal distance between the vegetation indices of weed and straw for each test criterion, e.g. the following
Figure 5. Intensity quotient of the 42 samples from a mulched plant.
IR 0.24 0.20 0.19 0.25 0.21 0.16
Material
Ripwort-plantain (upper side) Ripwort-plantain (lower side) Dandelion (upper side) Wheat straw (lower side) Wheat straw (upper side) Pea straw (upper side)
0.08 0.05 0.03 0.25 0.21 0.13
R
Group 1
0.24 0.20 0.21 0.24 0.22 0.20
IR 0.08 0.05 0.03 0.20 0.23 0.13
R
Group 2
0.24 0.20 0.21 0.23 0.22 0.21
IR 0.06 0.06 0.03 0.20 0.19 0.13
R
Group 3
0.25 0.20 0.21 0.21 0.25 0.22
IR 0.06 0.06 0.02 0.18 0.18 0.12
R
Group 4
TABLE I Raw values of infrared and red reflectance
0.24 0.20 0.19 0.16 0.24 0.22
IR 0.06 0.04 0.03 0.16 0.19 0.12
R
Group 5
0.22 0.19 0.18 0.12 0.21 0.20
IR
0.06 0.03 0.03 0.14 0.18 0.13
R
Group 6
0.21 0.18 0.17 0.10 0.21 0.20
IR
0.06 0.03 0.03 0.12 0.18 0.12
R
Group 7
514 H.-R. LANGNER ET AL.
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
515
Figure 6. Intensity differences of the 42 samples from a mulched plant.
Figure 7. NDVI values of the 42 samples from a mulched plant.
equation Q = Min(vegetationindexweed) − Max(vegetationindexstraw)
(5)
Table II lists the medians of the calculated results for all six materials and for all 4 vegetation indices. The medians of every column were calculated in the last line, qualitatively related to Equation (5).
516
H.-R. LANGNER ET AL.
TABLE II Vegetation indices and the quality parameter Q Material
Difference
Quotient
NDVI
DIRT
Ripwort-plantain (upper side) Ripwort-plantain (lower side) Dandelion (upper side) Wheat straw (lower side) Wheat straw (upper side) Pea straw (upper side) Quality parameter Q = Min (vi weed) − Max (vi straw)
0.16 0.15 0.16 0.00 0.03 0.08
1.05 1.00 1.06 0.63 0.64 0.74
0.57 0.60 0.73 0.00 0.08 0.24
0.57 0.60 0.73 0.00 −0.08 −0.10
0.07
0.26
0.33
0.57
Figure 8. DIRT values of the 42 samples from a mulched plant.
The quality parameter Q indicates the difference between the smallest vegetation index of weeds and the biggest vegetation index of straw materials. The largest difference arises in the column DIRT with 0.57, the newly defined vegetation index. DIRT therefore allows the unambiguous identification of weeds on mulched areas and was the most favourable criterion of the evaluation. The application of the DIRT brought a decision safety improved considerably in the test since the influence of straw and plant residue parts is suppressed. The swelling value β is a parameter whose value was set above the minimum of the red vegetable intensity. The value of parameter β was 0.12 during all tests carried out.
A SPECIAL VEGETATION INDEX FOR THE WEED DETECTION
517
5. Discussion For automatic weed ratings on mulched cropland and for sensor based herbicide application in PA a new vegetation index DIRT was derived and tested. The classification between weed seedlings and background (mulched cropland) was experimentally tested during weed ratings on a tractor. The criterion DIRT works well for the main step of weed image processing – the background discrimination, e.g. the decision between weed pixels or background pixels. The test results of the evaluation against three other vegetation indices show that the DIRT gives the highest decision quality. However this new index DIRT needs further investigations for weed detection on mulched areas or as decision criterion for other applications in sensor based precision agriculture. The decision quality using the DIRT is not only valuable for automatic weedratings on mulched cropland. The weed detection accuracy opens the possibility for a high weeding speed in practice. The higher working speed tends to lower the contrast of weed images. In practical applications a high decision quality can compensate slow contrasts in weed images. Another goal for using the DIRT vegetation index is to raise the working speed of sensor based weeding technologies to a level that is comparable with conventional weeding technologies.
References Aitkenhead, M. J., Dalgetty, I. A., Mullins, C. E., McDonald, A. J. S. and Strachan, N. J. C.: 2003, ‘Weed and crop discrimination using image analysis and artificial intelligence methods’, Comp. Electr. Agric. 39, 157–171. Baret, F. and Guyot, G.: 1991, ‘Potentials and limits of vegetation indices for LAI and PAR assessment’, Remote Sens. Environ. 35, 161–173. Biller, R. H. and Schicke R.: 2000, ‘Multi-frequency optical identification of different weeds and crops for herbicide reduction in precision agricukture’, In: Proc. 5th Intern. Conf. on Precision Agriculture, 16.-17.7.2000, University of Minnesota, St. Paul, MN, USA. Buschmann, C. and Nagel, E.: 1993, ‘In vivo spectroscopy and internal optics of leaves as basis for remote sensing of vegetation’, Int. J. Remote Sensing 14, 711–722. Chapron, M., Requena-Esteso, M., Boissard, P. and Assemat, L.: 1999, ‘A method for recognizing vegetal species from multispectral images’, Precision Agriculture 1999, 2nd European Conference on Precision Agriculture, Part 1, 239–248. Sheffield Academic Press, UK. Christensen S. and Heisel, T. 1998: ‘Patchspraying using historical, manual and real-time monitoring of weeds in cereals’, Z. Pflanzenkrankheiten Pflanzenschutz Sonderheft XVI, 257– 263. Christensen, S., Heisel, T. and Paice, M.: 1999, ‘Simulation of long term Alopecurus myosuroides population using three patchspraying strategies’, in: Precision Agriculture 1999, 2nd European Conference on Precision Agriculture, Part 2 (Stafford), pp. 977–987. Academic Press,Sheffield, UK. Felton, W. L. and McCloy, K. R.: 1992, ‘Spot spraying’, Agric. Eng. 11, 9–12. Gerhards, R., S¨okefeld, M., Nabout, A., Therburg, R. D. and K¨uhbauch, W.: 2002, ‘Online weed control using digital image analysis’, Z. Pflanzenkrankheiten Pflanzenschutz Sonderheft XVIII, 421–427.
518
H.-R. LANGNER ET AL.
Gitelson, A. A.: 2004, ‘Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation’, J. Plant Physiol. 161, 165–173. Heisel, T., Christensen, S. and Walter, A. M.: 1997, ‘Validation of weed patch spraying in spring barley – preliminary trial’, in: Precision Agriculture 1997, 1st European Conf. on Precision Agriculture, Vol 2: Technology, IT and Management (ed. Stafford), BIOS Scientific Publishers, Warwick, UK, pp. 879–886. Hock, B., Fedtke, C. and Schmidt, R. R.: 1995, Herbizide–Entwicklungen, Anwendungen, Wirkungen, Nebenwirkungen, Georg Thieme Verlag, Stuttgart, Germany. Johnson, G. A., Mortensen, D. A. and Martin, A. R.: 1995, ‘A simulation of herbicide use based on weed spatial distribution’, Weed Res. 35, 197–205. Kropff, M. J.: 1988, ‘Modelling the effects of weeds on crop production’, Weed Res. 8, 465–471. Lamb, D. W. and Brown, R. B.: 2001, ‘Remote sensing and mapping of weeds in crops’, J. Agric. Eng. Res. 78, 117–125. Langner, H. R., B¨ottger, H., Schmidt, H. and Wartenberg, G.: 2003, ‘Abschlussbericht zum Thema: Messsystem zur Bewertung des Unkrautvorkommens’, BMBF Teilprojekt 02(5444), 09/2001–10/2003. Laser2000: 2003, ‘Catalogue image processing products’, Comp. Laser2000, Weßling (Germany) Bulletin No. 1004, S.49–50. Lindquist, J. L., Dieleman, J. A., Mortensen, D. A., Johnson, G. A. and Wyse-Pester, D. Y.: 1998, ‘Economic importance of managing spatially heterogeneous weed populations’, Weed Technol. 12, 7–13. Marchant, J. A., Andersen, H. J. and Onyango, C. M.: 2001, ‘Evaluation of an imaging sensor for detecting vegetation using different waveband combinations’, Comp. Electr. Agric. 32, 101–117. Oppelt, N. and Mauser, W.: 2004, ‘Hyperspectral remote sensing – A tool for the derivation of plant nitrogen and its spatial variability within maize and wheat canopies’, Precision Agric. 493–498. Perez, A. J., Lopez, F., Benlloch, J. V. and Christensen, S.: 2000, ‘Colour and shape analysis techniques for weed detection in cereal fields’, Comp. Electr. Agric. 25, 197–212. Philipp, I., Nordmeyer, H. and Rath, T.: 2003, ‘Algorithms for discrimination of weeds in sugar beet fields using digital image processing’, Proceedings 4. European Conf. on Precision Agriculture, Programme book of ECPA – ECPLF, 527–528, Wageningen Academic Publishers. Philipp, I., Nordmeyer, H. and Rath T.: 2003, ‘Automatische und manuelle Unkrautkartierung in Zuckerr¨ubenbest¨anden im Vergleich’, Computerbildanalyse in der Landwirtschaft, Workshop 2003, Bornimer Agrartechnische Berichte, Heft 28, Institut f¨ur Agrartechnik Bornim. Rondeaux G., Steven M. and Baret, F.: 1996, ‘Optimisation of soil-adjusted vegetation indices’, Remote Sensing Environ. 55, 95–107. Tian, L.: 2002, ‘Development of a sensor-based precision herbicide application system’, Comp. Electr. Agric. 36, 133–149. Wartenberg, G. and Dammer, K. H.: 2001, ‘Site-specific real time application of herbicides in practice’, in: Grenier, G. und S. Blackmore (eds), 3rd European Conf. on Precision Agriculture, pp. 617–622. West, J. S., Bravo, C., Oberti, R., Lemaire, D., Moshou, D. and McCartney, H. A.: 2003, ‘The potential of optical canopy measurement for targeted control of field crop diseases’, Annu. Rev. Phytopathol. 41, 593–614.