Precision Agric (2007) 8:173–185

DOI 10.1007/s11119-007-9037-x

Development of an agricultural crops spectral library

and classification of crops at cultivar level using
hyperspectral data

N. Rama Rao Æ P. K. Garg Æ S. K. Ghosh

Published online: 1 August 2007

Ó Springer Science+Business Media, LLC 2007

Abstract In the context of a growing interest in remote sensing for precision agriculture
applications, the utility of space-borne hyperspectral imaging for the development of a
crop-specific spectral library and automatic identification and classification of three cul-
tivars for each of rice (Oryza sativa L.), chilli (Capsicum annuumL.), sugarcane (Sac-
charum officinarum L.) and cotton (Gossipium hirsutum L.) crops have been investigated in
this study. The classification of crops at cultivar level using two spectral libraries devel-
oped using hyperspectral reflectance data at canopy scale (in-situ hyperspectral measure-
ments) and at pixel scale (Hyperion data) has shown promising results with 86.5 and 88.8%
overall classification accuracy, respectively. This observation highlights the possible
integration of in-situ hyperspectral measurements with space-borne hyperspectral remote
sensing data for automatic identification and discrimination of various crop cultivars.
However, considerable spectral similarity is observed between cultivars of rice and sug-
arcane crops which may pose problems in the accurate identification of various crop
cultivars.

Keywords Spectral library
Remote sensing
Spectral matching
Hyperion

Classification

Introduction

Remote sensing applications in precision agriculture have been steadily increasing in

recent years due to improvements in spatial and spectral resolutions of remotely sensed

N. R. Rao
P. K. Garg
S. K. Ghosh

Centre for Remote Sensing, Department of Civil Engineering, Indian Institute of Technology,
Roorkee 247 667, India

N. R. Rao (&)
Department of Geospatial Sciences, RMIT University, Melbourne, VIC 3001, Australia
e-mail: [email protected]

123
174 Precision Agric (2007) 8:173–185

imagery. There is a rapidly growing interest in methods for automatic plant identification
in agricultural research. In particular, precision farming needs cost and time effective
methods for identification of various crop types in order to evaluate crop condition and
recommend suitable measures (Lutman and Perry 1999). In particular, information on the
type of crop cultivar is essential for proper implementation of precision farming methods.
Different crop cultivars of a single crop type may have variable growth and yield char-
acteristics owing to genetic potential, besides differences in the requirement and appli-
cation of inputs like manures, fertilizers and crop management practices. A key
requirement for precision agriculture is to identify various types of crop cultivars and
weeds automatically, so that fertilizers, herbicides and other agrochemicals could be
applied selectively (Robert 1999).
Traditionally, agricultural crops are identified using broadband satellite imagery by the
classification of satellite imagery with statistical classifiers such as the maximum likeli-
hood classifier (Philipp and Rath 2002). These statistical classifiers use training samples
which can be collected from various sources that include topographical maps, high-spatial
resolution satellite images and field visits. The limitation with this method, however,
becomes apparent when a map for a larger region comprised of multiple images is needed,
or for images acquired in different years. The analyst is faced with either needing training
sets from every scene or mosaicing the scenes together before any processing occurs. The
first option is not likely to be feasible and even defeats the purpose of using a remote
sensing approach. The other, that of mosaicing the images, was not practical due to large
file sizes and the computing resources necessary to process them together as a single file.
Furthermore, the success of species level crop classification using broadband satellite
imagery is very limited owing to the spectral similarity of various crop species and the finer
spectral differences among various crop species, which are not detectable in broadband
remote sensing data (Broge and Leblanc 2000). A method is needed to classify satellite
images for the same classes using training sets imported from one image, or using data
from a spectral library.
Recent advances in sensor technology have led to the development of hyperspectral
remote sensing imaging devices which can obtain high-spectral resolution radiance data for
each location (pixel) within the field of view. It is expected that such detailed spectral data
will permit the unique identification of most surface types of rocks, soils and vegetation,
provided that the spatial resolution of the data is sufficient to represent a single surface type
for each spectrum (Chen et al. 1999).
In the context of a growing interest in remote sensing for vegetation applications, the
possibilities of hyperspectral imaging for the extraction of information relevant to agri-
culture as well as vegetation demand detailed understanding of spectral signatures in terms
of position of feature specific absorption bands, shape of the spectrum, spectral variability
and similarity of various types of vegetation species. In particular, credible information on
correlation of field-collected spectral signatures of various crops with spectral reflectance
of space-borne hyperspectral sensors is important for the automatic identification and
quantification of various vegetation types in an area. One possible way of integrating the
in-situ spectral data with space-borne hyperspectral data is the availability of a spectral
library of various surface features of interest.
Theoretically, it is expected that if a comprehensive spectral library for specific agri-
cultural crop species is developed by taking into consideration various factors such as time
and year, growth stage, type of cultivar, soil and climate, it is possible to label the unknown
crop spectra by the method of spectral matching of signatures of crop species derived from
hyperspectral data (Price 1992). Further, if the spectral matching method provides labelling

123
Precision Agric (2007) 8:173–185 175

of unknown spectra correctly, this could be useful for automatic identification and
extraction of various crop species and may also lead to proper selection of training pixels
for mapping of various crop cultivars in an area. Price (1994) suggested the need to build
spectral libraries such that they can account for intra-species variability and unique dis-
crimination of various types of soil, mineral and vegetation species.
Several libraries of reflectance spectra of natural and man-made materials are available
for public use. These libraries provide a source of reference spectra that can aid the
interpretation of hyperspectral and multispectral images. Usually, these spectral libraries
are provided as a database add-in to current commercial software used for analysing
hyperspectral imagery. However, these spectral libraries have been developed and main-
tained for minerals, rocks and some vegetation species which are forest species and are not
for agricultural crops. Further, the spectral library may contain sufficiently large number of
reference spectra to accommodate variability of crop cultivar type, soil types and local
climate in order to meet the conditions required for spectral matching of unknown crop
spectra with the crop spectral library. Extensive literature review reveals that there has
been no systematic study carried out for the development of agriculture-specific spectral
libraries and their subsequent utility for the identification and classification of various crop
cultivars.
The objectives of this research were (1) development of cultivar based spectral library
for rice, sugarcane, chilli and cotton crops at mature growth phase using data from
Hyperion image and in-situ hyperspectral measurements and (2) to test the potential of the
developed spectral library for identification and classification of some important crop
cultivars of rice, sugarcane, chilli and cotton in Andhra Pradesh state, India. Further, this
study aims to provide a correlation of field based crop spectra to a set reference crop
spectra collected from Hyperion data.
The anticipated relevance of the work presented in this manuscript is to evaluate the
possibility of generating satellite based crop maps at cultivar level for precision farming
needs and to highlight the need of initiating further studies across the globe by various
researchers to bring out a comprehensive methodology suitable for operational use across
geographical locations. Further, this study is expected to be part of a regional level con-
ceptual system development for the electronic dissemination of information required by
precision producers.

Materials and methods

Study area

The study area is part of Guntur district, Andhra Pradesh state of India lying between
16°70 31@–16°500 55@N latitude and 79°400 37@–79°440 49@E longitude. A location map of the
study area is shown in Fig. 1. The annual rainfall of the district is 889 mm. The soils in the
area are fertile and are broadly classified as Black cotton, Red loamy and sandy loamy.
Black cotton area constitutes 70%, Red loamy 24% and sandy loamy 6% of the area. Most
of the cultivation in the study area is based on canal irrigation. The study has been carried
out for three different cultivars each of rice (cultivar names: BPT1768, MTU1001 and
MTU5293), sugarcane (cultivar names: CO6097, 85A261 and 84A125), chilli (cultivar
names: LCA235, LCA334 and LCA305) and cotton (cultivar names: LAHH4, Savitha and
LK861). The selection of these four crops and their cultivars is based on the fact that these
are dominant crops and cultivars in the study area during the kharif season (July to

123
176 Precision Agric (2007) 8:173–185

Fig. 1 Location map of the study area

November) and that the socio-economic condition of the local population depends con-
siderably on production and marketing price of these crops.

Data used

The data used in this study have been collected from several sources such as Hyperion data
for 28 September 2003, in-situ hyperspectral data using field spectroradiometer (GER
3700, Spectra Vista Corp., New York, NY, USA) and other ancillary data like crop type,
cultivar, date of sown/transplant, plant height, soil type etc. The Hyperion instrument on
the Earth Observing-1 (EO-1) satellite of NASA records visible light and other reflected
electromagnetic energy ranging from 0.4 to 2.5 mm in 220 channels that are 10-nm wide.
Its spatial resolution of 30 m and the orbit of the satellite complement those of Landsat.
Extensive field visits have been carried out to identify and locate large patches of
different cultivars of the selected crops which are at mature stage using stratified random
sampling method. The spatial location of these crop patches has been noted using DGPS
and overlaid on the Hyperion image. These locations were subsequently used as references
for spectral data extraction from the Hyperion image. Ancillary data such as date of crop
sowing, cultivar of crops has been collected from the local farmers.

Pre-processing of Hyperion image

The Hyperion image has been georeferenced with the help of Survey of India toposheets
(Scale 1: 250,000) and resampled to 30 m-pixel size using nearest neighbourhood method.
The image processing has been done using ERDAS IMAGINE-8.6 software. Since
Hyperion operates from a space platform with consequently modest surface signal levels
and full-column atmospheric effects, its data require careful processing. The striping has
been minimized using the Fast Fourier Transform technique using ERDAS IMAGINE
image processing software. The destriped Hyperion image is subjected to further

123
Precision Agric (2007) 8:173–185 177

processing for standard atmospheric corrections using FLAASH atmospheric correction

software module. All the pixels of Hyperion image belonging to non-vegetation, shrubs
and grasses classes, which are not part of classes of interest in the present study, have been
eliminated from the image by creating a mask using a threshold value of 0.23 for Nor-
malized Difference Vegetation Index (NDVI) image generated from Hyperion.

Methodology adopted

As indicated in the objectives of the study, three basic steps of analysis are required and are
shown in Fig. 2.

Building of spectral library using field spectroradiometer data (spectral library-1)

Spectral data on the selected cultivars of all the four crops were collected using spectro-
radiometer during the period 26–30 September 2003. The spectroradiometer collects data
in 704 narrow spectral bands with bandwidth ranging from 1.5 nm in the Visible-Near
Infra-Red (VNIR) to 8.5 nm in the Shortwave-Infra-Red (SWIR) wavelength. At each
measurement site in the field, 5–8 reflectance measurements were consistently taken along
a transect, with a nadir view from a height of 1.2 m for rice, cotton, chilli and 1.75 m for
sugarcane using a 18° FOV. Typical transect was 20–60 m in length with one spectra at
every 10 m interval. In this way, 200 samples each of three cultivars of the selected crops
(except sugarcane, for which 130 each of the three cultivars) at mature growth stage have
been collected. All the spectral measurements have been collected during 1100–1400 h
noontime to avoid the impact of illumination changes on the spectral responses. The
spectral data has been converted into reflectance with the help of readings taken

Fig. 2 Methodology flowchart Development of

spectral libraries

In-situ spectral
measurements Calibrated
GPS Hyperion image

Collection and Location of crop Collection of

calibration reflectance sample points reflectance values
values

Spectral module of ENVI- Spectral module of ENVI-

4.1 to give Spectral library-1 4.1 to give Spectral library-2

Classify Hyperion image using Spectral library-1 & Spectral library-

2

Classified images of crop varieties

Comparative analysis of accuracy of classified images

123
178 Precision Agric (2007) 8:173–185

simultaneously with a Barium Sulphate white reference panel. Spectral data were recorded
as reflectance values and then imported into a spreadsheet for further analyses. Individual
spectra were plotted and examined to ensure no erroneous values were incorporated in
subsequent analyses. The spectral library, which contains 2,190 reference spectra, has been
built using spectral module of ENVI 4.1 image processing software. The naming (label-
ling) of each spectrum of the spectral library has been done in such a way that the name of
the spectrum indicates name of the crop, its cultivar and sample number (indicator of site
of sample collected) of the cultivar. Further, the spectral library has been spectrally re-
sampled to similar wavebands of Hyperion image for meaningful comparison with
Hyperion imagery.

Building of spectral library using hyperion image (spectral library-2)

Based on the field visits, 2,190 samples of pixel-based retrieved reflectance spectra of
selected crop cultivars have been extracted from the calibrated Hyperion image. These
samples are selected from the same locations of the study area, at which the in-situ
hyperspectral measurements have been carried out. Care has been exercised to ensure that
there is no difference in the soil type, management practices and date of sowing/trans-
plantation for each crop cultivar. The spectral library, which contains 2,190 reference
spectra, has been built using the same procedure followed in the above section.

Classification and spectral matching of hyperion image with spectral library

Spectral matching and classification of various cultivars of the selected crops have been
carried out using a hyperspectral classification algorithm, viz, Spectral Angle Mapper
(SAM) (Kruse et al. 1993). The SAM is a physically based spectral classification algorithm
that uses an n-dimensional angle to match pixels to reference spectra. The algorithm
determines the spectral similarity between two spectra by calculating the angle between the
spectra, treating them as vectors in a space with dimensionality equal to the number of
bands. This technique, when used on calibrated reflectance data, is relatively insensitive to
illumination and albedo effects (Kruse et al. 1993).

Results and analysis

Classification of hyperion image using reference spectra from spectral library-1

The Hyperion image has been classified using SAM algorithm with the help of reference
spectral data from spectral library-1. The accuracy of the classified imagery is evaluated by
generating a contingency matrix using 1,388 independent reference samples, which were
not part of the reference samples used for the development of the spectral library (Table 1).
The resulting classified image is shown in Fig. 3. The overall accuracy of the classified
map is 86.5%, which indicates the potential of hyperpsectral imagery for cultivar level
classification of crops. The cotton crop showed highest average classification accuracy
(89.6%) followed by the chilli crop (87.4%), whereas the rice and sugarcane crops were
classified with average classification accuracies of 83.9 and 85.1%, respectively, which
were low among all the four crops.

123
 Table 1 Contingency matrix showing accuracy assessment of classified image using reference spectra from in-situ hyperspectral measurements (Spectral Library-1)
Actual reference pixels

Variety name Cotton Rice Sugarcane Chillies Total

Classified imagery pixels LAHH4 Savitha LK861 BPT1768 MTU1001 MTU5293 CO6907 85A261 84A125 LCA235 LCA334 LCA305

Cotton LAHH4 103 9 4 2 1 116

Savitha 12 104 5 121
LK861 4 2 104 110
Precision Agric (2007) 8:173–185

Rice BPT1768 99 1 2 13 115

MTU1001 2 109 13 1 125
MTU5293 13 11 97 121
Sugarcane CO6907 86 6 8 100
85A261 7 5 3 91 5 111
84A125 4 1 9 87 101
Chillies LCA235 107 3 6 116
LCA334 9 101 9 119
LCA305 6 11 114 131
Total 119 115 113 121 126 116 103 107 100 124 115 129 Overall
Producer’s accuracy (%) 86.5 90.4 92.0 81.8 86.5 83.6 83.4 85.0 87.0 86.2 87.8 88.3 accuracy:
86.5%
179

123
180 Precision Agric (2007) 8:173–185

Fig. 3 Classified map showing different cultivars of rice, sugarcane, cotton and chilli crops (Image
classification has been done using reference spectra from spectral library built using in-situ hyperspectral
measurements)

Examination of the contingency matrix (Table 1) reveals a considerable spread in the

classified pixels of rice and sugarcane crops. Many of the rice crop pixels, particularly
those belong to rice crop cultivar BPT1768; have been misclassified as sugarcane crop
pixels. Similarly, some of the sugarcane crop pixels have been misclassified as rice crop
pixels indicating a possible spectral similarity of rice and sugarcane crop cultivars. Indeed,
there has been considerable spectral similarity observed between rice and sugarcane crops,
which might have contributed to the mis-identification of some of the rice pixels as
sugarcane and vice-versa (Fig. 4).
Another issue which needed to be examined is per-class classification accuracy. The
cotton crop cultivar LK861 has been classified with highest accuracy (92%), while rice
crop cultivar BPT1768 has been classified with lowest classified accuracy among all the
four crops cultivars studied. This may be partly due to the spectral variability expected
within a single crop cultivar and the possible spectral similarity expected among various

123
Precision Agric (2007) 8:173–185 181

Fig. 4 Spectral profiles of agricultural crops selected in the present study

Fig. 5 Spectral profiles showing variability of spectral reflectance among three cultivars of cotton crop

cultivars of the rice crop, which is evident from the spread of classified pixels among
different cultivars of all the four crops. Further, it is evident that the spread in the wrongly
classified pixels of all the four crops is more pronounced between two cultivars of the same
crop. For example, out of the 119 reference pixels of cotton crop cultivar LAHH4,
103 pixels have been classified correctly and out of the 16 wrongly classified pixels,
12 pixels have been misclassified as Savitha and 4 pixels as LK861 cotton cultivars
indicating a close relationship between LAHH4 and Savitha cotton cultivars.
In order to find out the possible cause of mixing of these two cultivars, the spectral
profiles of all the three cotton cultivars are plotted (Fig. 5). It has been observed that the
cultivars LAHH4 and Savitha show similar spectral shape and magnitude, which may be
attributed to the identification of LAHH4 as Savitha and vice-versa (Fig. 5). Contrary to

123
182 Precision Agric (2007) 8:173–185

this, the LK861 showed clear differences in the shape of the spectrum, particularly, in the
NIR region indicating prospects of fair discrimination from the other two cultivars of
cotton. A similar type of spectral relationship is observed among different cultivars of the
remaining three crops.
However, the overall accuracy of the classified map (86.5%) shows a good correlation
between in-situ spectral data and space-borne hyperspectral data. Thus, it can be inferred
that well calibrated space-borne hyperspectral data can provide fair level of discrimination
amongst agricultural crops at the cultivar level.

Classification hyperion image using reference spectra from spectral library-2

In order to know the performance of reference spectra derived from in-situ spectral
measurements for the classification of space-borne hyperspectral data, the classification
results obtained from reference spectra collected from in-situ spectral measurements are
compared with classification results obtained with reference spectra collected from
Hyperion image. The Hyperion imagery has been classified with the help of reference
spectra from spectral library-2 (i.e. reference spectra was from the same Hyperion image).
The resulting classified map and accuracy assessment matrix is shown in Fig. 6 and
Table 2.
The classification of various cultivars of rice, sugarcane, chilli and cotton crops using
reference spectra from a Hyperion image has given interesting per-class classification
accuracy with an overall classification accuracy of 88.8% (Table 2). The comparatively
higher overall accuracy (88.8%) indicates the potential of hyperspectral imagery for cul-
tivar level classification of agricultural crops, at least for crops studied in this article.
Further, the trend of per-class classification is similar to that obtained with reference
spectra from in-situ spectral measurements. In this case also, the cotton crop has been
classified with highest average classification accuracy (91.1%) followed by chilli crop
(89.4%). However, there was a 4% increase in the average classification accuracy for the
rice crop (87.8%) while using reference spectra from the Hyperion image. There was very
marginal increase in the average classification accuracy of sugarcane crop while using
reference spectra from the Hyperion image (86.7%) over that obtained with reference
spectra from in-situ spectral measurements (85.1%).
Further, there was a marginal difference of 2.3% overall accuracy between the two
classified images generated using reference spectra obtained from ground based observa-
tion and Hyperion image, with Hyperion data having higher accuracy. This can be
explained by the closeness of the amplitude of the Hyperion imagery derived reference
spectra to the reflectance values of the Hyperion image. Further, discrepancy may also
arise due to some residual in atmospheric corrections in the Hyperion image and slight
illumination changes which may have occurred while collecting in-situ spectral
measurements.

Discussion

The present study has established the possibility of discrimination in one particular area of
agricultural crops at cultivar level by spectral matching of a crop-specific spectral library
developed using in-situ hyperspectral measurements with space-borne hyperspectral
reflectance data (Hyperion). The overall classification accuracy obtained in this study,

123
Precision Agric (2007) 8:173–185 183

Fig. 6 Classified map showing different cultivars of rice, sugarcane, cotton and chilli crops (Image
classification has been done using reference spectra from spectral library built using Hyperion image)

particularly, when using reference spectra from a Hyperion image (88.8%) encourages the
application of this methodology to other sites and crop types and suggests the use of
Hyperion images for species level discrimination of various agricultural crops. Further,
there is scope to improve the accuracy of crop area estimates by incorporating the spectral
variability observed in hyperspectral signatures of various crop cultivars in the image
classification process.
However, this study has confined itself to the discrimination of cultivars of four agri-
cultural crops in a single geographical region and at mature growth phase, at which the soil
background reflectance problem is minimal. Further studies are suggested to validate the
results of this study by considering multiple locations and at various growth phases to
ensure the general applicability of the approach proposed in this research. The develop-
ment of crop-specific spectral libraries is tedious and care should be taken to accommodate
the largest spectral and spatial variability expressed by crops under natural conditions.
Caution should be exercised while extending the results of this study to other locations as
the dynamic nature of crops and highly variable weather and soil conditions may pose

123
 Table 2 Contingency matrix showing accuracy assessment of classified image using reference spectra from Hyperion image (Spectral Library-2)
184

Actual reference pixels

Variety name Cotton Rice Sugarcane Chillies Total

123
Classified imagery pixels LAHH4 Savitha LK861 BPT1768 MTU1001 MTU5293 CO6907 85A261 84A125 LCA235 LCA334 LCA305

Cotton LAHH4 107 7 2 1 4 121

Savitha 7 102 4 2 115
LK861 5 6 107 118
Rice BPT1768 104 2 5 6 117
MTU1001 6 114 10 2 132
MTU5293 8 6 101 115
Sugar cane CO6907 89 6 5 100
85A261 3 4 5 94 9 115
84A125 3 5 86 94
Chillies LCA235 111 3 3 117
LCA334 4 105 9 118
LCA305 8 5 113 126
Total 119 115 113 121 126 116 103 107 100 124 115 129 Overall
Producer’s accuracy (%) 89.9 88.6 94.6 85.9 90.4 87.0 86.4 87.8 86.0 89.5 91.3 87.5 accuracy:
88.8%
Precision Agric (2007) 8:173–185
Precision Agric (2007) 8:173–185 185

problems in using spectral libraries developed in one region to discriminate the same crops
at other locations.

Conclusions

This study was conducted using natural field grown crop samples at canopy scale to pixel
scale to determine whether it would be possible theoretically to separate various crop
cultivars based solely on crop reflectance with the help of spectral libraries. The dis-
crimination of as many as 12 crop cultivars belonging to rice, sugarcane, chilli and cotton
crops in a semi-arid part of India has shown promising results with 86.5 and 88.8% spectral
variability explained by hyperspectral reflectance data at canopy scale and at pixel scale.
This observation highlights the possible use of hyperspectral remote sensing data for
development of vegetation, specifically, crop-specific spectral libraries for identification
and mapping of various crop cultivars. Of the four crops studied, cotton crop cultivars have
been discriminated with highest accuracy (88.6–94.6%) followed by chilli crop cultivars
(87.5–91.3%) whereas considerable spectral similarity was observed between LAHH4
cultivars of rice and sugarcane crops that may be detrimental to the unique identification
and automatic classification of agricultural crop species. Further, this study supports the
idea of integration of in-situ hyperspectral measurements with space-borne Hyperspectral
imagery for better calibration and improvement in the discrimination of various classes of
interest.

Acknowledgements The first author would like to acknowledge Dr. N.V. Krishna Rao, Managing
Director, m/s SVS Agro-Biotech Research Services Pvt. Ltd., Andhra Pradesh, India for his encouragement
and for partial financial assistance. Thanks are also due to Dr. N.S. Rao for his assistance during the
collection of in-situ hyperspectral measurements and to All India Council for Technical Education, Govt. of
India, New Delhi for the financial support received to in the form of National Doctoral Fellowship. Authors
gratefully acknowledge the three anonymous reviewers for their extremely constructive criticism and for
directions given for improving quality of this manuscript.

References

Broge NH, Leblanc E (2000) Comparing prediction power and stability of broadband and hyper-
spectral vegetation indices for estimation of green leaf area index and canopy chlorophyll
density. Remote Sens Environ 76:156–172
Chen JM, Leblanc SG, Miller JR, Freemantle J, Loechel SE,Walthall CL, Innanen KA, White HP
(1999) Compact airborne spectrographic imager (CASI) used for mapping biophysical
parameters of boreal forests. J Geophys Res 104:27945–27958
Kruse FA, Lefkoff AB, Dietz JB (1993) Expert system-based mineral mapping in northern Death
Valley, California/Nevada using the airborne visible/infrared imaging spectrometer (AVIRIS).
Remote Sens Environ 44:309–336
Lutman PJW, Perry NH (1999) Methods of weed patch detection in cereal crops. In: Weeds, vols
1–3. Brighton Conference, British Crop Protection Council, UK, pp 627–634
Philipp I, Rath T (2002) Improving plant discrimination in image processing by use of different
colour space transformations. Comput Electron Agric 35(1):1–15
Price JC (1992) Variability of high resolution crop reflectance spectra. Int J Remote Sens 14:
2593–2610
Price JC (1994) How unique are spectral signatures. Remote Sens Environ 49:181–186
Robert PC (1999) Precision agriculture: research needs and status in the USA. In: Stafford JV (ed)
Proceedings of the 2nd European conference on Precision Agriculture, Sheffield Academic
Press, Sheffield, UK, pp 19–33

123