Review Received: 3 August 2012
Revised: 23 October 2013
Accepted article published: 6 November 2013
Published online in Wiley Online Library: 6 December 2013
(wileyonlinelibrary.com) DOI 10.1002/ps.3677
Potential use of ground-based sensor technologies for weed detection Gerassimos G. Peteinatos,∗ Martin Weis, Dionisio Andujar, Victor Rueda ´ Ayala and Roland Gerhards Abstract Site-specific weed management is the part of precision agriculture (PA) that tries to effectively control weed infestations with the least economical and environmental burdens. This can be achieved with the aid of ground-based or near-range sensors in combination with decision rules and precise application technologies. Near-range sensor technologies, developed for mounting on a vehicle, have been emerging for PA applications during the last three decades. These technologies focus on identifying plants and measuring their physiological status with the aid of their spectral and morphological characteristics. Cameras, spectrometers, fluorometers and distance sensors are the most prominent sensors for PA applications. The objective of this article is to describe-ground based sensors that have the potential to be used for weed detection and measurement of weed infestation level. An overview of current sensor systems is presented, describing their concepts, results that have been achieved, already utilized commercial systems and problems that persist. A perspective for the development of these sensors is given. c 2013 Society of Chemical Industry Keywords: Site Specific Weed Management; Weed Detection; Optical Sensors; Distance Sensors
1
INTRODUCTION
190
Nowadays, agricultural production needs to provide increasingly higher yields while complying to stricter rules regarding environmental aspects. Site-specific weed management (SSWM) can aid in fulfilling these economical and environmental aspects.1,2 Mostly, SSWM focuses on spraying weed patches only (nozzles on/off) or doing variable-rate spraying according to weed species composition, distribution and density.3,4 This is achieved with the aid of weed maps or on-line during the application.5 Weed maps are generated from pre-treatment sampling or historical maps from preceding years.6 Weed populations in agricultural fields have been found to be spatially and temporally variable within and between agricultural fields.7,8 They depend on site characteristics, such as soil type, but are mostly influenced by human interactions during crop management, including tillage, crop rotation and weed control methods.9 Weed populations are often stable within fields over time, due to persistent weed seed bank and vegetative propagules in the soil.10 They produce and drop seeds mostly before the crop is harvested, explaining the high spatial stability of weed populations. Until now, the in-field heterogeneity, even though researched, in practice has received little attention concerning weed control methods. Herbicides and mechanical weed control treatments are mostly applied uniformly across whole fields regardless of the variation in weed densities and weed species composition. Taking into account the weed variability within a field contributes to a reduction of herbicide use, while simultaneously maintaining high weed control efficacy.5,8,11 – 13 Spatial variation in weed populations must therefore be taken into account for weed control decisions. Pest Manag Sci 2014; 70: 190–199
A major step for the introduction of SSWM into practice is the automation of weed detection with sensors. Different spatial resolutions of measurements can be achieved with aerial imaging and ground-based sensors. Aerial imaging with the aid of airborne methods (satellites, aircraft, unmanned aerial vehicles) can identify within-field variations of coverage and detect large, already developed weed patches.14 – 18 Medlin et al. were able to achieve above 70% accuracy in identifying broad-leaved weeds in a soybean field (Glycine max), with a spatial resolution of 1 m per pixel if plant density was above 10 plants m− 2 .19 Weed patches were also detected by Bajwa and Tian with 4.5–5.3 m per pixel, but they failed to classify their composition.20 Nevertheless, aerial imaging – especially satellite-based systems – rely on the weather conditions at the time of measurement. Weed applications usually need to be done when weeds are at a early growing stage. Therefore the time window between taking measurements, deriving conclusions concerning weed composition and coverage, making decisions regarding the necessity of an application and the actual application is short. Overcast and cloudy sky during the measurement time window could irreparably delay the weed management.3,21 Another problem arises from the similarity of spectra which can be retrieved from completely different canopy mixtures.3,22,23 In addition, even if there are some weed species like Equisetum arvense L. that can be identified from the air, in
∗
Correspondence to: G. G. Peteinatos, Department of Weed Science, University of Hohenheim, Otto-Sander-Str. 5, 70599 Stuttgart, Germany. E-mail: [email protected] Department of Weed Science, University of Hohenheim, 70599 Stuttgart, Germany
www.soci.org
c 2013 Society of Chemical Industry
Ground-based sensor technologies for weed detection most cases weed identification, especially at early growth stages, requires high spatial resolutions which are not yet provided by aerial systems.24,25 Ground-based systems are relatively independent of the weather conditions and could potentially be used for determining weed density and identifying weed species. A series of sensor technologies have been developed. Most of them have presented promising results, which can be used for in-field operation, but in most cases there are no commercial products yet. Weed detection can be achieved either by directly recognizing weed species in the field, or indirectly by measuring total plant coverage, leaf area index, photosynthetic activity or plant height.26 In the latter case, values of the field parameters are interpreted as weed-free crop biomass, or weed infestation ratio, based on a priori knowledge. The objective of this article is to describe ground-based sensor technologies which directly target weed–crop discrimination or can be used to measure the weed infestation indirectly. Some recommendations for the combination of sensor and application technologies to improve weed detection strategies are provided.
2
NON-IMAGING SENSORS
Non-imaging sensors yield measurements of a spot in the field (footprint). The following non-imaging weed-sensing approaches can be distinguished: first, identification of plants with the aid of their spectral characteristics such as reflected or emitted light (spectrometers, fluorescence sensors); second, characteristics such as the height above ground (light detection and ranging (LiDAR), ultrasonic sensors).
Pest Manag Sci 2014; 70: 190–199
The spectral properties of different weed and crop species have been explored and used for their discrimination under controlled conditions, which cannot easily be transferred to field conditions.23,25,39,40 Since the ambient light in field conditions can change rapidly due to general weather conditions like clouds and atmospheric variances, the measurements have to be rectified according to the daytime and yearly cycles of the sun illumination (sun angle). There are even influences because of the changing angle between the measurement instrument and the illumination source. For near-range measurements the influence of the atmosphere can be omitted due to the possibility of a reference measurement of the ambient light conditions (building a ratio of reference and measurement) or repeated calibration in the field.41 A more detailed insight concerning the calibration of a spectrometric system is given in Noble et al.25 Furthermore, the limited spatial information limits the use of spectral systems for multi-species differentiation. Spectral identification approaches are complex and need proper prior information (spectral signatures) of the study objects, which are not available under the conditions found in the field. In chemometrics – the science of extracting information from chemical systems – the identification and quantification of various chemical substances are quite efficient, due to the existence of spectral libraries. Such libraries are not publicly available for weed identification. Therefore, weed–crop or weed–weed classifications needs to take into account more information. A combination of the spectral characteristics with other information, like shape or coverage from imaging, is usually necessary.40 Additionally, the output of spectrometric sensors depends on the mixture of all objects in their field of view. Therefore, small weeds, which can be detected from an imaging sensor with its higher spatial resolution, may pass undetected and unrecognized.42 Commercial spectral sensors for precision agriculture (PA) in general have already been on the market for over two decades. Most of them detect certain spectral properties of the crop reflectance. They are usually applied for nitrogen application and for determining the physiological status of the plants. Weed detection with these sensors has also been a research topic, yet the current results are far from a commercial product.43 Yara N-sensor (Yara International ASA, Norway) measures a series of different spectra to determine indices related to the nitrogen status. Based on the raw measurement data (if accessible), other spectral indices can be calculated (e.g. NDVI). The first variant of the sensor was a passive sensor, reliant on natural ambient light to function, but the newest versions have an artificial light source (xenon flash). 2.2 Optoelectronic sensors Sensors that focus on very few (usually one or two), quite specific spectral bands are called optoelectronic sensors. They focus mostly on bands in the red/near-infrared (R/NIR) spectrum. Non-imaging optoelectronic sensors can discriminate between plant presence and absence by measuring indices correlated with plant coverage values.44 Although they cannot discriminate between species (crops and weeds), they can be useful, for example, in detecting weeds between rows in row crops. Commercial sensors of this type measure reflectance properties in the NIR and R wavelengths to derive an index comparable to the NDVI. Felton and McCloy developed the spot-spraying system DetectSpray for real-time weed control based on an optoelectronic sensor placed in front of the nozzle, but even though it showed some promising results
c 2013 Society of Chemical Industry
wileyonlinelibrary.com/journal/ps
191
2.1 Spectrometric sensors Spectrometers measure the reflection intensities at wavelengths within a range of the electromagnetic spectrum, ranging from ultraviolet (UV), visible light, to near-infrared (NIR). The spectral resolution depends on the sensor and up to 2 nm resolution has been achieved (FieldSpec 4 Hi-Res Spectroradiometer, ASD Inc., Boulder, CO, USA). Each plant absorbs certain parts of the ambient light due to photosynthetically active compounds or other lightabsorbing pigments in the leaves.17,25,27 – 30 For weed detection, the most common sensors rely on measuring a spectrum covering the visible and NIR light and calculating vegetation indices such as the normalized difference vegetation index (NDVI), which can be easily derived from spectral data.31 They correlate well with general plant coverage, leaf area index and biomass. NDVI is often used for nutrition prescriptions, but tends to saturate for high biomass values.32 – 34 Spectrometers can provide information to differentiate plants from soil, although they are mostly not able to differentiate different species of plants. In fact, green leaves are characterized by a high reflectance in the green and NIR and a low reflectance in the red and blue spectra. On the other hand, the reflectance curve of bare soil shows a linear increase of reflectance from the blue to the NIR light.31,35 – 38 This can be seen in Fig. 1: the reflectance curves for plants and soil differ clearly, but a distinction of different plant species is difficult to achieve. Additionally, the spectral response of a plant changes during different growth stages. The spectrometer, having a certain measurement field of view, integrates over all curves within the field-of-view area, and therefore the received signal is a mixture of different amounts of plant species and soil.
www.soci.org
www.soci.org this system is no longer commercially available.36,45 Dammer and Wartenberg also proposed a similar system by emitting light at 650 nm and 830 nm and measuring an NIR/R quotient.46 At the early development stages the measurements are highly correlated with plant frequency, while at later development stages it correlates with plant coverage. The relation of coverage and weed density is also dependent on the height structure of the plants, since large plants contribute most of the coverage in heterogeneous situations.47 WeedSeeker GreenSeeker (Trimble Agriculture, Sunnyvale, CA, USA), WEEDit (gps-Ag Pty Ltd, Kangaroo Flat, Australia) and Crop Circle ACS-470 (Holland Scientific Inc., Lincoln, NE, USA) are products in this market.48 – 50 All the above are active sensors, providing an illumination of their own. These products are usually combined with a sprayer, which is automatically turned on when the NDVI difference of NIR to R reflectance exceeds a specified threshold, indicating a high vegetation cover in the field of view of the sensor. Andujar ´ et al. showed the possibilities of WeedSeeker for weed treatments in maize crops.21 In three fields they compared weed maps created with an array of three WeedSeeker units and digital images. They reported an agreement above 78% between the two spray maps. Most of the reported errors concern underestimation of weed coverage. Other researchers have also reported sufficient weed detection with the use of WeedSeeker.51,52
ge Green Yellow (G) (Y)
Red (R)
Near infrared (NIR)
0.2
Ultra violet Blue (UV) (B)
Red ed
0.4
0.6
Triticum aestivum Lamium purpureum Triticum spelta Papaver rhoeas Soil
The sensors Multiplex (Force A, France), PAM fluorometry (Heinz Walz GmbH, Germany) and MiniVegN (Fritzmeier Umwelttechnik GmbH & Co. KG, Germany) are all active sensors that measure chlorophyll fluorescence. Multiplex and PAM sensors can radiate in a series of different wavelengths (UV, blue, green, red for Multiplex and blue, red, NIR for PAM sensors), measuring BGF fluorescence and CLa fluorescence. While PAM sensors need some time for dark adaptation of the plants, therefore making it technically challenging to transfer this technology to the field, Multiplex is capable of measuring fluorescence without dark adaptation. These measurements can be used for early detection of plant stress, nitrogen status, ripening stage, chlorophyll content and herbicide resistance. With high spectral resolution, fluorescence has already been used to quantify phenotypes.55 UV-induced chlorophyll fluorescence has also been applied to discriminate plant species based on their characteristic leaf structure.53,56,57 Longchamps et al. measured a range of fluorescence spectra of maize, grass weeds and broad-leaved weeds on greenhouse conditions under natural illumination.58 They used UV radiation centred at 327 nm to measure fluorescence induced around 400 nm. They classified the three plant species groups based on their distinct spectral signatures with a recognition rate above 90%. Tyystj¨arvi etal., with the use of a PAM sensor, developed a method called fluorescence fingerprinting, in which leaves are exposed to a series of different colours and intensities of light to record changes in the fluorescence of CLa.59 The emitted light curve could be used to identify plant species with an accuracy of more than 90% under laboratory conditions. Later, Tyystj¨arvi et al. applied a similar approach under field conditions and achieved 90% recognition of maize and weeds, when the measured plants were shaded for 1 s before measuring.57 The MiniVegN sensor has been used in greenhouse studies and in the field to map the spatial distribution of weed species in arable crops.60 Red and far-red CLa fluorescence was induced by a red laser. When the laser hit a plant, fluorescence was induced and recorded. If the sensor moved with a steady speed, plant density was highly correlated with the number of plant hits received per second. Due to the high frequency of measurements (500 Hz), the sensor was able to achieve a high spatial resolution and a coefficient of determination R2 of above 0.9 was achieved for the correlation between weed density and number of hits. Gerhards etal. examined a winter wheat field in the year 2011 with Alopecurus myosuroides Huds. as the dominant weed species and reported
0.0
Reflectance (white reference = 1)
2.3 Fluorescence sensors After exposing plants with radiation (for a specific amount of time and intensity) leaves emit fluorescent radiation. The fluorescent wavelength is longer than the excitation light wavelength, and emitted in a small amount of time after receiving light. The intensity of fluorescence highly depends on leaf properties and physiological state.53 Plants emit fluorescent light, mostly due to the chemical compounds of chlorophyll, polyphenol and or flavonol anthocyanins. UV radiation stimulates blue-green fluorescence (BGF) (450 nm) of the leaves, mostly due to flavonols and anthocyanins located in the epidermis of leaves. Chlorophyll a (CLa) and b (CLb) emits fluorescence in the red to far-red range of the spectrum (680–700 and 735–750 nm).54 Invivo CLb transfers all its excitation to CLa; therefore in the field measurable fluorescence originates mainly from CLa. Fluorescence induced by UV or laser light highly correlates with leaf properties and the physiological status of the plants. Different species produce different amplitudes of BGF. The ratio of BGF to CLa fluorescence is highly related on the plant species and can be used for plant identification.53
GG Peteinatos et al.
350
400
450
500
550
600
650 700 750 Wavelength [nm]
800
850
900
950
1000
192
Figure 1. Reflectance spectra in the visual and NIR wavebands. Plant reflectance exhibits maxima for green and infrared, with a steep increase from red to infrared. Soil reflectance is linear in this spectral range.
wileyonlinelibrary.com/journal/ps
c 2013 Society of Chemical Industry
Pest Manag Sci 2014; 70: 190–199
Ground-based sensor technologies for weed detection 75% correct identification compared to manual weed counting.60 Fluorescence has provided promising results for the detection of various agronomic characteristics. The fingerprinting method proposed by Tyystj¨arvi et al. in combination with the weed density measurements proposed by Gerhards et al. is a weed-mapping approach that still needs to be evaluated.59,60 2.4 Distance sensors Apart from using some kind of electromagnetic wavelength, some sensors measure the distance from the sensor. Distance sensors can be used for estimating the plant height and to derive biomass values.61 – 64 Two major sensing technologies are available in this category: light detection and ranging (LiDAR) sensors and ultrasonic sensors. LiDAR sensor technology measures the distance to the target with a laser beam. This can be done using one of the following methods:2,65,66 • by measuring the time needed for the laser pulse to travel between the transmitter and the receiver of the sensor, reflected by the target (time-of-flight LiDAR); • by measuring the phase difference produced from the emitted laser beam and the reflected one (phase-shift measurement LiDAR).
Pest Manag Sci 2014; 70: 190–199
the plants in a side view. Binary images could be created from consecutive lines of measurements and perpendicular, horizontal vehicle movements. Ultrasonic sensors use an ultrasonic pulse to measure the distance to the target, based on the time-of-flight technique.63 Ultrasonic sensors have been used for weed detection and discrimination.63,71,72 The last echo can be attributed to the soil, while previous echoes belong to crop matter.62 Andujar et al. detected weeds in the inter-row space of maize, and discriminated between grasses and broad-leaved weeds with an ultrasonic sensor.63 If monocots and dicots are at the same height, the output of the ultrasonic sensor is different due to the different amplitude of the generated echo, since broad-leaf plants have a larger echoing surface. They obtained more than 77% accuracy classifying weeds into the two groups. The method seems promising in row crops, if the major weed species have a significant, measurable height. In the case of cereal crops, the ultrasonic sensor can provide high correlations with the weed biomass only at the early stages of crop development, while it cannot locate weeds overlapped by crop plants.71,72 However, in some specific cases the reflected echo cannot be differentiated and sensor fusion with spectral or imaging sensors could solve the problem. The ultrasonic sensor can only provide one distance value, whenever the minimum surface reflecting the sound back to the sensor exceeds a certain threshold. If the measured surface has a significant crop coverage and the crop is higher than the weeds, the measurement from the ultrasonic reflects the crop coverage rather than the weeds under the crop.
3
IMAGING SENSORS
3.1 Imaging sensors Weed detection with the aid of imaging sensors is the most investigated technique applied.21,73 Efforts to implement machine vision for agricultural purposes have existed for a significant amount of time. Portable imaging and analysis technology in agricultural research was introduced over three decades ago. The wide availability of RGB sensors and their relative low costs are reasons to apply them in-field for measurements. To differentiate plants from soil, (additional) infrared wavebands are better suited due to the high infrared reflectance of plants. The NIR/R ratio is higher than for RGB wavelengths.74 Changes of filters led to recent developments of NDVI cameras based on standard RGB camera technology.75,76 Many researchers have applied machine vision systems for the detection of weeds in agricultural fields.13,21,77 – 81 Some of the developed systems cannot just identify weed patches based on the fact that weed-infested spots contain more biomass than non-infected ones, but can also differentiate weeds from crops and among different weed species with varying degrees of success.40,74,78,82,83 There are commonly several steps involved in the process. Most systems include digital image acquisition and image segmentation (subdividing images into regions of similar characteristics), then extracting plant shape, colour, texture and location features. ¨ Sokefeld and associates proposed a bi-spectral camera, taking images of red (610–690 nm) and NIR wavelengths (>700 nm and
Revised: 23 October 2013
Accepted article published: 6 November 2013
Published online in Wiley Online Library: 6 December 2013
(wileyonlinelibrary.com) DOI 10.1002/ps.3677
Potential use of ground-based sensor technologies for weed detection Gerassimos G. Peteinatos,∗ Martin Weis, Dionisio Andujar, Victor Rueda ´ Ayala and Roland Gerhards Abstract Site-specific weed management is the part of precision agriculture (PA) that tries to effectively control weed infestations with the least economical and environmental burdens. This can be achieved with the aid of ground-based or near-range sensors in combination with decision rules and precise application technologies. Near-range sensor technologies, developed for mounting on a vehicle, have been emerging for PA applications during the last three decades. These technologies focus on identifying plants and measuring their physiological status with the aid of their spectral and morphological characteristics. Cameras, spectrometers, fluorometers and distance sensors are the most prominent sensors for PA applications. The objective of this article is to describe-ground based sensors that have the potential to be used for weed detection and measurement of weed infestation level. An overview of current sensor systems is presented, describing their concepts, results that have been achieved, already utilized commercial systems and problems that persist. A perspective for the development of these sensors is given. c 2013 Society of Chemical Industry Keywords: Site Specific Weed Management; Weed Detection; Optical Sensors; Distance Sensors
1
INTRODUCTION
190
Nowadays, agricultural production needs to provide increasingly higher yields while complying to stricter rules regarding environmental aspects. Site-specific weed management (SSWM) can aid in fulfilling these economical and environmental aspects.1,2 Mostly, SSWM focuses on spraying weed patches only (nozzles on/off) or doing variable-rate spraying according to weed species composition, distribution and density.3,4 This is achieved with the aid of weed maps or on-line during the application.5 Weed maps are generated from pre-treatment sampling or historical maps from preceding years.6 Weed populations in agricultural fields have been found to be spatially and temporally variable within and between agricultural fields.7,8 They depend on site characteristics, such as soil type, but are mostly influenced by human interactions during crop management, including tillage, crop rotation and weed control methods.9 Weed populations are often stable within fields over time, due to persistent weed seed bank and vegetative propagules in the soil.10 They produce and drop seeds mostly before the crop is harvested, explaining the high spatial stability of weed populations. Until now, the in-field heterogeneity, even though researched, in practice has received little attention concerning weed control methods. Herbicides and mechanical weed control treatments are mostly applied uniformly across whole fields regardless of the variation in weed densities and weed species composition. Taking into account the weed variability within a field contributes to a reduction of herbicide use, while simultaneously maintaining high weed control efficacy.5,8,11 – 13 Spatial variation in weed populations must therefore be taken into account for weed control decisions. Pest Manag Sci 2014; 70: 190–199
A major step for the introduction of SSWM into practice is the automation of weed detection with sensors. Different spatial resolutions of measurements can be achieved with aerial imaging and ground-based sensors. Aerial imaging with the aid of airborne methods (satellites, aircraft, unmanned aerial vehicles) can identify within-field variations of coverage and detect large, already developed weed patches.14 – 18 Medlin et al. were able to achieve above 70% accuracy in identifying broad-leaved weeds in a soybean field (Glycine max), with a spatial resolution of 1 m per pixel if plant density was above 10 plants m− 2 .19 Weed patches were also detected by Bajwa and Tian with 4.5–5.3 m per pixel, but they failed to classify their composition.20 Nevertheless, aerial imaging – especially satellite-based systems – rely on the weather conditions at the time of measurement. Weed applications usually need to be done when weeds are at a early growing stage. Therefore the time window between taking measurements, deriving conclusions concerning weed composition and coverage, making decisions regarding the necessity of an application and the actual application is short. Overcast and cloudy sky during the measurement time window could irreparably delay the weed management.3,21 Another problem arises from the similarity of spectra which can be retrieved from completely different canopy mixtures.3,22,23 In addition, even if there are some weed species like Equisetum arvense L. that can be identified from the air, in
∗
Correspondence to: G. G. Peteinatos, Department of Weed Science, University of Hohenheim, Otto-Sander-Str. 5, 70599 Stuttgart, Germany. E-mail: [email protected] Department of Weed Science, University of Hohenheim, 70599 Stuttgart, Germany
www.soci.org
c 2013 Society of Chemical Industry
Ground-based sensor technologies for weed detection most cases weed identification, especially at early growth stages, requires high spatial resolutions which are not yet provided by aerial systems.24,25 Ground-based systems are relatively independent of the weather conditions and could potentially be used for determining weed density and identifying weed species. A series of sensor technologies have been developed. Most of them have presented promising results, which can be used for in-field operation, but in most cases there are no commercial products yet. Weed detection can be achieved either by directly recognizing weed species in the field, or indirectly by measuring total plant coverage, leaf area index, photosynthetic activity or plant height.26 In the latter case, values of the field parameters are interpreted as weed-free crop biomass, or weed infestation ratio, based on a priori knowledge. The objective of this article is to describe ground-based sensor technologies which directly target weed–crop discrimination or can be used to measure the weed infestation indirectly. Some recommendations for the combination of sensor and application technologies to improve weed detection strategies are provided.
2
NON-IMAGING SENSORS
Non-imaging sensors yield measurements of a spot in the field (footprint). The following non-imaging weed-sensing approaches can be distinguished: first, identification of plants with the aid of their spectral characteristics such as reflected or emitted light (spectrometers, fluorescence sensors); second, characteristics such as the height above ground (light detection and ranging (LiDAR), ultrasonic sensors).
Pest Manag Sci 2014; 70: 190–199
The spectral properties of different weed and crop species have been explored and used for their discrimination under controlled conditions, which cannot easily be transferred to field conditions.23,25,39,40 Since the ambient light in field conditions can change rapidly due to general weather conditions like clouds and atmospheric variances, the measurements have to be rectified according to the daytime and yearly cycles of the sun illumination (sun angle). There are even influences because of the changing angle between the measurement instrument and the illumination source. For near-range measurements the influence of the atmosphere can be omitted due to the possibility of a reference measurement of the ambient light conditions (building a ratio of reference and measurement) or repeated calibration in the field.41 A more detailed insight concerning the calibration of a spectrometric system is given in Noble et al.25 Furthermore, the limited spatial information limits the use of spectral systems for multi-species differentiation. Spectral identification approaches are complex and need proper prior information (spectral signatures) of the study objects, which are not available under the conditions found in the field. In chemometrics – the science of extracting information from chemical systems – the identification and quantification of various chemical substances are quite efficient, due to the existence of spectral libraries. Such libraries are not publicly available for weed identification. Therefore, weed–crop or weed–weed classifications needs to take into account more information. A combination of the spectral characteristics with other information, like shape or coverage from imaging, is usually necessary.40 Additionally, the output of spectrometric sensors depends on the mixture of all objects in their field of view. Therefore, small weeds, which can be detected from an imaging sensor with its higher spatial resolution, may pass undetected and unrecognized.42 Commercial spectral sensors for precision agriculture (PA) in general have already been on the market for over two decades. Most of them detect certain spectral properties of the crop reflectance. They are usually applied for nitrogen application and for determining the physiological status of the plants. Weed detection with these sensors has also been a research topic, yet the current results are far from a commercial product.43 Yara N-sensor (Yara International ASA, Norway) measures a series of different spectra to determine indices related to the nitrogen status. Based on the raw measurement data (if accessible), other spectral indices can be calculated (e.g. NDVI). The first variant of the sensor was a passive sensor, reliant on natural ambient light to function, but the newest versions have an artificial light source (xenon flash). 2.2 Optoelectronic sensors Sensors that focus on very few (usually one or two), quite specific spectral bands are called optoelectronic sensors. They focus mostly on bands in the red/near-infrared (R/NIR) spectrum. Non-imaging optoelectronic sensors can discriminate between plant presence and absence by measuring indices correlated with plant coverage values.44 Although they cannot discriminate between species (crops and weeds), they can be useful, for example, in detecting weeds between rows in row crops. Commercial sensors of this type measure reflectance properties in the NIR and R wavelengths to derive an index comparable to the NDVI. Felton and McCloy developed the spot-spraying system DetectSpray for real-time weed control based on an optoelectronic sensor placed in front of the nozzle, but even though it showed some promising results
c 2013 Society of Chemical Industry
wileyonlinelibrary.com/journal/ps
191
2.1 Spectrometric sensors Spectrometers measure the reflection intensities at wavelengths within a range of the electromagnetic spectrum, ranging from ultraviolet (UV), visible light, to near-infrared (NIR). The spectral resolution depends on the sensor and up to 2 nm resolution has been achieved (FieldSpec 4 Hi-Res Spectroradiometer, ASD Inc., Boulder, CO, USA). Each plant absorbs certain parts of the ambient light due to photosynthetically active compounds or other lightabsorbing pigments in the leaves.17,25,27 – 30 For weed detection, the most common sensors rely on measuring a spectrum covering the visible and NIR light and calculating vegetation indices such as the normalized difference vegetation index (NDVI), which can be easily derived from spectral data.31 They correlate well with general plant coverage, leaf area index and biomass. NDVI is often used for nutrition prescriptions, but tends to saturate for high biomass values.32 – 34 Spectrometers can provide information to differentiate plants from soil, although they are mostly not able to differentiate different species of plants. In fact, green leaves are characterized by a high reflectance in the green and NIR and a low reflectance in the red and blue spectra. On the other hand, the reflectance curve of bare soil shows a linear increase of reflectance from the blue to the NIR light.31,35 – 38 This can be seen in Fig. 1: the reflectance curves for plants and soil differ clearly, but a distinction of different plant species is difficult to achieve. Additionally, the spectral response of a plant changes during different growth stages. The spectrometer, having a certain measurement field of view, integrates over all curves within the field-of-view area, and therefore the received signal is a mixture of different amounts of plant species and soil.
www.soci.org
www.soci.org this system is no longer commercially available.36,45 Dammer and Wartenberg also proposed a similar system by emitting light at 650 nm and 830 nm and measuring an NIR/R quotient.46 At the early development stages the measurements are highly correlated with plant frequency, while at later development stages it correlates with plant coverage. The relation of coverage and weed density is also dependent on the height structure of the plants, since large plants contribute most of the coverage in heterogeneous situations.47 WeedSeeker GreenSeeker (Trimble Agriculture, Sunnyvale, CA, USA), WEEDit (gps-Ag Pty Ltd, Kangaroo Flat, Australia) and Crop Circle ACS-470 (Holland Scientific Inc., Lincoln, NE, USA) are products in this market.48 – 50 All the above are active sensors, providing an illumination of their own. These products are usually combined with a sprayer, which is automatically turned on when the NDVI difference of NIR to R reflectance exceeds a specified threshold, indicating a high vegetation cover in the field of view of the sensor. Andujar ´ et al. showed the possibilities of WeedSeeker for weed treatments in maize crops.21 In three fields they compared weed maps created with an array of three WeedSeeker units and digital images. They reported an agreement above 78% between the two spray maps. Most of the reported errors concern underestimation of weed coverage. Other researchers have also reported sufficient weed detection with the use of WeedSeeker.51,52
ge Green Yellow (G) (Y)
Red (R)
Near infrared (NIR)
0.2
Ultra violet Blue (UV) (B)
Red ed
0.4
0.6
Triticum aestivum Lamium purpureum Triticum spelta Papaver rhoeas Soil
The sensors Multiplex (Force A, France), PAM fluorometry (Heinz Walz GmbH, Germany) and MiniVegN (Fritzmeier Umwelttechnik GmbH & Co. KG, Germany) are all active sensors that measure chlorophyll fluorescence. Multiplex and PAM sensors can radiate in a series of different wavelengths (UV, blue, green, red for Multiplex and blue, red, NIR for PAM sensors), measuring BGF fluorescence and CLa fluorescence. While PAM sensors need some time for dark adaptation of the plants, therefore making it technically challenging to transfer this technology to the field, Multiplex is capable of measuring fluorescence without dark adaptation. These measurements can be used for early detection of plant stress, nitrogen status, ripening stage, chlorophyll content and herbicide resistance. With high spectral resolution, fluorescence has already been used to quantify phenotypes.55 UV-induced chlorophyll fluorescence has also been applied to discriminate plant species based on their characteristic leaf structure.53,56,57 Longchamps et al. measured a range of fluorescence spectra of maize, grass weeds and broad-leaved weeds on greenhouse conditions under natural illumination.58 They used UV radiation centred at 327 nm to measure fluorescence induced around 400 nm. They classified the three plant species groups based on their distinct spectral signatures with a recognition rate above 90%. Tyystj¨arvi etal., with the use of a PAM sensor, developed a method called fluorescence fingerprinting, in which leaves are exposed to a series of different colours and intensities of light to record changes in the fluorescence of CLa.59 The emitted light curve could be used to identify plant species with an accuracy of more than 90% under laboratory conditions. Later, Tyystj¨arvi et al. applied a similar approach under field conditions and achieved 90% recognition of maize and weeds, when the measured plants were shaded for 1 s before measuring.57 The MiniVegN sensor has been used in greenhouse studies and in the field to map the spatial distribution of weed species in arable crops.60 Red and far-red CLa fluorescence was induced by a red laser. When the laser hit a plant, fluorescence was induced and recorded. If the sensor moved with a steady speed, plant density was highly correlated with the number of plant hits received per second. Due to the high frequency of measurements (500 Hz), the sensor was able to achieve a high spatial resolution and a coefficient of determination R2 of above 0.9 was achieved for the correlation between weed density and number of hits. Gerhards etal. examined a winter wheat field in the year 2011 with Alopecurus myosuroides Huds. as the dominant weed species and reported
0.0
Reflectance (white reference = 1)
2.3 Fluorescence sensors After exposing plants with radiation (for a specific amount of time and intensity) leaves emit fluorescent radiation. The fluorescent wavelength is longer than the excitation light wavelength, and emitted in a small amount of time after receiving light. The intensity of fluorescence highly depends on leaf properties and physiological state.53 Plants emit fluorescent light, mostly due to the chemical compounds of chlorophyll, polyphenol and or flavonol anthocyanins. UV radiation stimulates blue-green fluorescence (BGF) (450 nm) of the leaves, mostly due to flavonols and anthocyanins located in the epidermis of leaves. Chlorophyll a (CLa) and b (CLb) emits fluorescence in the red to far-red range of the spectrum (680–700 and 735–750 nm).54 Invivo CLb transfers all its excitation to CLa; therefore in the field measurable fluorescence originates mainly from CLa. Fluorescence induced by UV or laser light highly correlates with leaf properties and the physiological status of the plants. Different species produce different amplitudes of BGF. The ratio of BGF to CLa fluorescence is highly related on the plant species and can be used for plant identification.53
GG Peteinatos et al.
350
400
450
500
550
600
650 700 750 Wavelength [nm]
800
850
900
950
1000
192
Figure 1. Reflectance spectra in the visual and NIR wavebands. Plant reflectance exhibits maxima for green and infrared, with a steep increase from red to infrared. Soil reflectance is linear in this spectral range.
wileyonlinelibrary.com/journal/ps
c 2013 Society of Chemical Industry
Pest Manag Sci 2014; 70: 190–199
Ground-based sensor technologies for weed detection 75% correct identification compared to manual weed counting.60 Fluorescence has provided promising results for the detection of various agronomic characteristics. The fingerprinting method proposed by Tyystj¨arvi et al. in combination with the weed density measurements proposed by Gerhards et al. is a weed-mapping approach that still needs to be evaluated.59,60 2.4 Distance sensors Apart from using some kind of electromagnetic wavelength, some sensors measure the distance from the sensor. Distance sensors can be used for estimating the plant height and to derive biomass values.61 – 64 Two major sensing technologies are available in this category: light detection and ranging (LiDAR) sensors and ultrasonic sensors. LiDAR sensor technology measures the distance to the target with a laser beam. This can be done using one of the following methods:2,65,66 • by measuring the time needed for the laser pulse to travel between the transmitter and the receiver of the sensor, reflected by the target (time-of-flight LiDAR); • by measuring the phase difference produced from the emitted laser beam and the reflected one (phase-shift measurement LiDAR).
Pest Manag Sci 2014; 70: 190–199
the plants in a side view. Binary images could be created from consecutive lines of measurements and perpendicular, horizontal vehicle movements. Ultrasonic sensors use an ultrasonic pulse to measure the distance to the target, based on the time-of-flight technique.63 Ultrasonic sensors have been used for weed detection and discrimination.63,71,72 The last echo can be attributed to the soil, while previous echoes belong to crop matter.62 Andujar et al. detected weeds in the inter-row space of maize, and discriminated between grasses and broad-leaved weeds with an ultrasonic sensor.63 If monocots and dicots are at the same height, the output of the ultrasonic sensor is different due to the different amplitude of the generated echo, since broad-leaf plants have a larger echoing surface. They obtained more than 77% accuracy classifying weeds into the two groups. The method seems promising in row crops, if the major weed species have a significant, measurable height. In the case of cereal crops, the ultrasonic sensor can provide high correlations with the weed biomass only at the early stages of crop development, while it cannot locate weeds overlapped by crop plants.71,72 However, in some specific cases the reflected echo cannot be differentiated and sensor fusion with spectral or imaging sensors could solve the problem. The ultrasonic sensor can only provide one distance value, whenever the minimum surface reflecting the sound back to the sensor exceeds a certain threshold. If the measured surface has a significant crop coverage and the crop is higher than the weeds, the measurement from the ultrasonic reflects the crop coverage rather than the weeds under the crop.
3
IMAGING SENSORS
3.1 Imaging sensors Weed detection with the aid of imaging sensors is the most investigated technique applied.21,73 Efforts to implement machine vision for agricultural purposes have existed for a significant amount of time. Portable imaging and analysis technology in agricultural research was introduced over three decades ago. The wide availability of RGB sensors and their relative low costs are reasons to apply them in-field for measurements. To differentiate plants from soil, (additional) infrared wavebands are better suited due to the high infrared reflectance of plants. The NIR/R ratio is higher than for RGB wavelengths.74 Changes of filters led to recent developments of NDVI cameras based on standard RGB camera technology.75,76 Many researchers have applied machine vision systems for the detection of weeds in agricultural fields.13,21,77 – 81 Some of the developed systems cannot just identify weed patches based on the fact that weed-infested spots contain more biomass than non-infected ones, but can also differentiate weeds from crops and among different weed species with varying degrees of success.40,74,78,82,83 There are commonly several steps involved in the process. Most systems include digital image acquisition and image segmentation (subdividing images into regions of similar characteristics), then extracting plant shape, colour, texture and location features. ¨ Sokefeld and associates proposed a bi-spectral camera, taking images of red (610–690 nm) and NIR wavelengths (>700 nm and
