Estimation of Nitrogen Content by Spectral Responses of Cabbage by knu24191


									An ASAE/CSAE Meeting Presentation                                                                        Paper Number: 041082

          Estimation of Nitrogen Content by Spectral Responses
           of Cabbage Seedlings Using Artificial Neural Network

                              C. T. Chen, Graduate Student; S. Chen, Professor
           Dept. of Bio-Industrial Mechatronics Engineering, National Taiwan University, Taipei,

                                           K. W. Hsieh, Associate Professor
            Dept. of Bio-Industrial Mechatronics Engineering, National Chung Hsing University,
                                             Taichung Taiwan.

           H. C. Yang; Assistant Researcher; S. Hsiao, I C. Yang, Graduate Student
           Dept. of Bio-Industrial Mechatronics Engineering, National Taiwan University, Taipei,

                                   Written for presentation at the
                           2004 ASAE/CSAE Annual International Meeting
                                    Sponsored by ASAE/CSAE
                      Fairmont Chateau Laurier, The Westin, Government Centre
                                     Ottawa, Ontario, Canada
                                         1 - 4 August 2004
Abstract. Reflectance spectra of leaves are informative for non-destructive monitoring of the
nutrition status. Nitrogen content of cabbage seedling leaves cultivated with five different fertilization
conditions was measured by Kjeldahl method, and was correlated with spectrum reflectance. Since
the wavelength of the band-pass filters for silicon CCD multi-spectral imaging was shorter than
1000nm, a selected band (450-950 nm) was considered in addition to the full band (400-2500 nm).
Step-wise multi-linear regression (SMLR) and artificial neural network (ANN) were used to evaluate
the effectiveness of the wavelength in determination of nitrogen content. The analytical results of
SMLR calibration equations with four significant wavelengths (566, 574, 1396, and 1530 nm;
rc2=0.82, SEC=9.85 mg/g, and SEV=12.35 mg/g) were significantly improved by an ANN model
(rc2=0.89, SEC=8.27 mg/g, and SEV=8.84 mg/g) with cross-learning and random- sampling
strategies, in which the over-fitting was greatly reduced. For developing a practical multi-spectral
The authors are solely responsible for the content of this technical presentation. The technical presentation does not necessarily reflect the
official position of ASAE or CSAE, and its printing and distribution does not constitute an endorsement of views which may be expressed.
Technical presentations are not subject to the formal peer review process, therefore, they are not to be presented as refereed publications.
Citation of this work should state that it is from an ASAE/CSAE meeting paper. EXAMPLE: Author's Last Name, Initials. 2004. Title of
Presentation. ASAE/CSAE Meeting Paper No. 04xxxx. St. Joseph, Mich.: ASAE. For information about securing permission to reprint or
reproduce a technical presentation, please contact ASAE at or 269-429-0300 (2950 Niles Road, St. Joseph, MI 49085-9659
imaging system with commercially available band-pass filters, the ANN model with four inputs of
490, 570, 600, and 680 nm was trained to obtain a comparable result (rc2=0.87, SEC=8.73 mg/g,
rv2=0.84, and SEV=9.60 mg/g) to the best calibration equation of SMLR with seven wavelengths in
the full band. The potential of using the ANN model developed in this study with multi-spectral
imaging to monitor the nitrogen content of cabbage seedling is expected.
Keywords. Artificial neural network, Cross-learning, Nitrogen content, Reflectance spectra

(The ASAE/CSAE disclaimer is on a footer on this page, and will show in Print Preview or Page
Layout view.)

The authors are solely responsible for the content of this technical presentation. The technical presentation does not necessarily reflect the
official position of ASAE or CSAE, and its printing and distribution does not constitute an endorsement of views which may be expressed.
Technical presentations are not subject to the formal peer review process, therefore, they are not to be presented as refereed publications.
Citation of this work should state that it is from an ASAE/CSAE meeting paper. EXAMPLE: Author's Last Name, Initials. 2004. Title of
Presentation. ASAE/CSAE Meeting Paper No. 04xxxx. St. Joseph, Mich.: ASAE. For information about securing permission to reprint or
reproduce a technical presentation, please contact ASAE at or 269-429-0300 (2950 Niles Road, St. Joseph, MI 49085-9659
Nitrogen content is crucial for the growth of plants; direct and non-destructive remote sensing of
their nutrient status is promising especially in cultivating the crop (Goel et al., 2003; Boegh et
al., 2002; Diker and Bausch, 2003). Nitrogen contents of plants in the form of dry powder, fresh
leaves and even the canopy have been estimated by calibrating the reflectance spectra with the
results of wet-chemical analysis (White et al., 2000; Grossman et al., 1996). Typically, step-wise
multi-linear regression (SMLR) and modified partial least square regression (MPLSR) were
applied to analyzed the spectra obtained with a spectrophotometer or a fiber–optic spectrometer
(Yoder and Pettigrew-Crosby, 1995; Serrano et al., 2002; García-Ciudad et al., 1999; Johnson
and Billow, 1996; Gillon et al., 1999).
SMLR method requires fewer parameters for a satisfactory fitting of calibration equation. Yorder
and Pettigrew (1995) successfully estimated the nitrogen contents in maple leaves using
photometric data of 3 and 5 different wavelengths, and the determination coefficients (r2) of
SMLR were 0.71 and 0.85, respectively. Compared with SMLR, MPLSR method is even better
for chemometric applications (Bolster et al., 1996; Gillon et al., 1999), but spectral data of full
(400-2500 nm) or partial (1100-2500 nm) bands and therefore a sophisticated optical system
with grating mechanism or tunable filters are demanded. The powerful calibration method is
therefore inappropriate for remote sensing studies using simple or field-type optical devices
such as photodiodes (Stone et al., 1996). Moreover, remote sensing including multi-spectral
imaging techniques (Goel et al., 2003; Wood et al., 2003) using simple optics devices provide
additional flexibility in eliminating environmental noise that will disturb the chemometric
prediction. Therefore, calibration methods using fewer wavelengths (i.e. SMLR) are more
suitable for general remote sensing purposes.
Wavelength determination is important for the cost reduction and precision of a remote sensing
system, but cares should be taken for the non-linear phenomena at the chosen wavelengths. To
calibrate a non-linear system, artificial neural network (ANN) with sigmoid transfer function and
error back-propagating training scheme was widely used (Liu et al., 2001; Hsieh et al., 2001;
Burks et al., 2000).
Recently, Hsieh et al., (2002) had developed an ANN classification model for carcass (94%
accuracy), but a large number of wavelengths was required. With fewer parameters, Mutanga
and Skidmore (2004) compared the ANN and SMLR to map the grass nitrogen concentration in
an African savanna rangeland. The ANN results showed a higher correlation coefficient (r=
0.96) for the training set, but the results for the test sets was not satisfactory (r= 0.77). In the
present approach for appraising the application potential of multi-spectral imaging system for in-
situ monitoring the nutrient status of cabbage seedlings, the significant wavelengths were
determined by SMLR method, and a new ANN training scheme was proposed for the nitrogen
content prediction. The scheme required fewer input nodes for simulating the spectral data
extracted from multi-spectral images of the seedlings, and a higher prediction accuracy was

Materials and methods

Cultivation of Cabbages Seedlings
Cabbage seedlings were planted in plastics pots of 8 cm diameter filled with medium composing
of peat moss and perlite (1:1); each pot was irrigated with 50 ml of water every two days, and

grown in 25℃ (day) and 20℃ (night) in a phytotron located in National Taiwan University. After
20 days of the conditioned growing, each pot was fertilized with 50 ml of HYPONeX No. 5
(HYPONeX, N: P: K=30:10:10) diluted to different extents (0, 500, 1000, 2000 and 4000 ppm).
The fertilized seedlings were continuously grown until the 25th day.

Spectra Acquisitions and Nitrogen Content Measurement
After 25 days of cultivation, the foliages (432 from 130 seedlings) were cut and put into the
small ring cup (NIRS 6500, FOSS NIRSystems) to measure the reflectance absorbance spectra
(Table 1). Total nitrogen contents of 118 randomly selected foliages were measured by Kjeldahl
method (García-Ciudad et al., 1999), and the results (TKN values, Table 1) were used to
calibrate with different spectra types (smooth, derivative, full and partial bands of raw data) by
MPLSR (WinISI software, FOSS NIRSystems). The best MPLSR equation, resulting from
comparing rc2, SEC (standard error of calibration) and SECV (standard error of cross
validation), was used to predict the nitrogen content of other samples.

Determination of multi-spectral imaging filters by SMLR
Spectra and nitrogen content of leaves (432 samples) were divided into the calibration set (324
samples) and validation set (108 samples) for SMLR analysis; the samples in both sets were
with similar nitrogen content distributions (Table 1). Raw spectra, from 400-2498 nm with 2 nm
resolution, were indicated as (400-2500, 2) and used to establish calibration equations with one
to seven significant wavelengths. By comparing r2, SEC, and SEV (standard error of
validation), the best equation was selected with understanding the capability of error
convergence by SMLR.
In order to extract the significant wavelengths for determining the filters of multi-spectral imaging
system, the spectra in the sensitive band of CCD camera from 450 nm to 950 nm at 2 nm
interval, indicated as (450-950, 2), were also analyzed by SMLR. Since the full width half
maximum (FWHM) of commercial band-pass filters are usually 10 or 20 nm, the spectra in 450-
950 nm were reconstructed by 10-points (20 nm) smoothing and selecting data points at 10 nm
interval. The reconstructed spectra were indicated as (450-950, 10) and used to analyze the
significant wavelengths to compare the results of the spectra types (400-2500, 2) and (450-950,
The significant wavelengths sets determined by SMLR in three types of spectra ((400-2500, 2);
(450-950, 2); (450-950, 10)) were then used to model the ANN for estimating nitrogen content of
leaves to obtain better calibration and prediction.

Conventional Modeling of ANN
The frame of adopted ANN was a perception multilayer network with error back-propagation as
the training scheme and the generalized delta rule for weighting and threshold values adjusting
(Fig. 1). The input nodes were the wavelengths determined by SMLR; the numbers of neuron
nodes in hidden layer 1 and 2 with the sigmoid transfer function were analyzed from 10 to 25 at
5 interval, and the singular output node with linear transfer function was used to calculate the
nitrogen content of leaves. Matlab 6.1 (The Mathworks, Inc.) was used to model the ANN with
adjustable learning rate and momentums to converge error quickly at each batch of training.
The initial values of weightings and thresholds were set consistent to escape the danger of
being trapped in local minima during a gradient-based search (Boger, 2003). The training
scheme of ANN used the calibration sample set divided from all samples as the main objects to
decrease error of the model and the remaining samples were as the validating objects to

confirm the situation of over-fitting. Besides, the Savebest strategy (Hsieh et al., 2002) of ANN
was also adopted. Therefore, the same sample sets of calibration and validation with SMLR
were analyzed to determine the proper number of nodes in hidden layers and to confirm the
ability of ANN for converging the error of nitrogen content prediction.
                Weightings              Weightings
                of Neuron    Sigmoid    of Neuron    Sigmoid                              Liner
                                                                of Neuron
                in hidden    transfer   in hidden    transfer                           transfer
                                                                 in output
  Spectral        layer 1    function    layer 2                                        function
                                                     function     layer 2                             Nitrogen
                Thresholds              Thresholds
                 of Neuron               of Neuron
                                                                 of Neuron
                 in hidden               in hidden
                                                                 in output
                   layer 1                layer 2
                                                                  layer 2

 Figure 1. The structure of ANN model with one input layer, two hidden layers, and one output
                     layer for traditional back-propagation training scheme.

Proposed Modeling of ANN
The proposed cross-learning schemes (Fig. 2) divided the samples of calibration into three
subsets. Every two subsets were combined as the learning set in the next step, and each
learning set was used as the calibration set in ANN modeling by turns. During the training
process, the weighting and threshold values were adjusted by the errors resulting from different
learning sets. The model of ANN with cross-learning scheme also validated by validation set
and used the Savebest strategy. Finally, based on the conception of cross-learning, partial
samples of calibration set were randomly selected into the learning set of ANN to calculate the
error for adjusting the weightings and thresholds. The error convergence values of all calibration
set were computed and compared with the results of validation set using the mentioned
strategy. Therefore, the results of three ANN types were compared at every training epoch for
their SEC and SEV values and at the best epoch for the rc2 and rv2 values.

                                                         Learning            If N=3×i
                                                          Set 1
                                Set 1

                                                         Learning            If N=3×i+1
             Calibration       Sample                     Set 2
              Sample            Set 2

                                                         Learning            If N=3×i+2
                                Set 3                     Set 3

   Figure 2. The new scheme of ANN training with cross-learning mechanism detailed in the
                                  experimental section.


Nitrogen contents in seedling leaves
In Table 1, the total nitrogen content of randomly selected samples (n= 118) were determined
by Kjeldahl method (TKN values), and the statistical results were tabulated. These analytical
results were used to develop MPLSR calibration equations for the reflectance absorbance
spectra of the samples. The spectra (1100-2500 nm) were smoothed (ten point smoothing) and
its second derivative spectra (data interval = 10) was taken to obtain the best calibration
equation by MPLSR. The best calibration results were rc2=0.97, SEC=4.24 mg/g, SECV=5.23
mg/g. The statistical results of the nitrogen contents obtained by Kjeldahl method and predicted
by MPLSR are similar, which indicates the homogeneity of the random sampling process.
In order to determine the significant wavelengths for selecting suitable filters of multi-spectral
imaging, leaves (432 samples) were divided into calibration (324 samples) and validation sets
(108 samples), the distributions of nitrogen contents were similar as revealed by MPLSR
prediction (Table 1). The samples were then analyzed by SMLR to determine the significant
Table 1. Statistical results of the nitrogen contents (mg/g) of seedling leaves for different
experimental groups
                        Number of     Mean    S.D.              Max.           Min.
TKN results of selected 118           40.85   25.46             88.43          3.47
MPLSR results of all    432           39.86   22.57             85.25          2.07
MPLSR results for       324           39.1    23.65             85.25          2.07
calibration set*
MPLSR results for       108           38.71   23.48             83.85          3.91
validation set*
* The sample sets were used for SMLR and ANN analysis.

Number of Signification Wavelengths and SMLR Error in Full Band
Table 2 shows the SMLR prediction errors obtained by analyzing one to seven significant
wavelengths in full band. The errors were converged until the seventh terms without over-fitting.
With seven wavelength calibration, the high linearity and low prediction errors can be achieved
without pretreatments of the raw spectra.
Table 2. SMLR results of one to seven wavelengths in full band.
                          Number of Wavelengths
                          1     2       3       4            5         6        7
Calibration set SEC       13.47 11.17 10.69 9.85             8.86      8.68     8.06

                 rc 2     0.67     0.77     0.79    0.82     0.86      0.86     0.88
Validation set   SEV      13.34    11.78   11.08    12.35    10.92     10.66    9.93

The reported remote sensing with multiple radiometers or spectral imaging system were less
than four bands (Stone et al., 1996; Diker and Bausch, 2003; Kostrzewski et al., 2003); four
terms were therefore considered to be the maximum for establishing the remote sensing model

with multi-spectral apparatus. However, as shown in Table 2, the results are not satisfactory,
and other modeling methods should be attempted to increase the accuracy. In the following,
filter sets of four and seven different wavelengths were determined from their SMLR results
calibrated for the spectra types ((400-2500, 2); (450-950, 2); (450-950, 10)) mentioned in the
experimental section.

Determining Filters for Multi-Spectral Imaging System
Table 3 listed the SMLR regression results of three types of spectra against four and seven
selected wavelengths. The calibration equation with 7 wavelengths from full spectral band was
the best, but the four wavelengths (1396 nm, 1536 nm, 1868 nm, and 2190 nm) in the near-
infrared region require advanced but expensive imaging devices such as InGaAs or HgCdTe
camera to construct a multi-spectral imaging system (Ungar et al., 2003). The applications will
be restricted and inappropriate for agricultural purposes. The significant wavelengths ranged
from 450 to950 nm are considered to be suitable for general and routine usage.
Although slightly less accurate than the full band result, the results of CCD sensing band of four
and seven wavelength calibration are acceptable with additional economic benefits for
assembling a CCD-based multi-spectral imaging system. The suitable wavelengths were
restricted further by commercially available band-pass filters, and calibration equation with
spectral data at 490, 570, 600, and 680 nm were found to be satisfactory for quantification
purposes. With respect to the last calibration results of the spectra of wider bandwidth (20 nm),
the data smoothing pretreatment may help in noise removal and thus improve the prediction
Table 3. Significant wavelengths and the prediction errors obtained by SMLR analysis.
Source spectra     Significant wavelengths                       SEC       rc2        SEV
(400-2500, 2)      566, 574, 1396, 1530                          9.85      0.82       12.35
                   566, 596, 686, 1396, 1536, 1868, 2190         8.06      0.88       9.93
(450-950, 2)       532, 540, 606, 686                            10.11     0.81       12.69
                   486, 574, ,596, 612, 630, 654, 692            9.30      0.84       11.10
(450-950, 10)      490, 570, 600, 680                            9.96      0.82       10.99
                   490, 570, 590, 640, 660, 670, 680             9.48      0.84       10.80

Error convergence of ANN
ANN with traditional training scheme was performed by dividing all samples into calibration and
validation sets, and the data of 4 significant wavelengths from full spectral band were input to
analyze the proper network structure of layer and node numbers. Figure 3A shows the training
results with one hidden layer, the error converged after 150 training epochs in case of 10 and 15
nodes, but the SEC and SEV show some inconsistency in case of the structure with higher node
number. Over-fitting probably occurred in the training course. Similar phenomena were more
obvious in the ANN model with two hidden layers (Fig. 3B). Both results revealed a dilemma of
the accuracy and consistence between calibration and validation sets, a training scheme based
on cross-learning was therefore designed.
The proposed training scheme with three alternating fixed learning sets was analyzed with the
same structures as the traditional scheme, the results were compared in Fig. 3C and Fig. 3D.
The SEV after 1000 training epochs converged to about 12 mg/g with one hidden layer and 10
to 25 nodes (Fig. 3C); over-fitting was reduced but not to a satisfactory extent. The results with

two hidden layers (Fig. 3D) also shows a reduced divergence in contrast to traditional ANN, but
the problematic over-fitting and inconsistent phenomena still existed.
Figure 3E and 3F are the results obtained with a similar cross-learning training scheme except
that the samples in the learning set were randomly selected from the calibration set. The
improvements obtained with every ANN structure are obvious. Regarding the error convergence
ability and the training cost, network of two hidden layers with 15 nodes was used in the

                               A                                             B

                C                                                                D

                    E                                         F

 Figure 3. The error convergence of ANN model consisting 4 inputs from full band with 10 (N10)
 to 25 nodes (N25): (A) one hidden and (B) two hidden layers by traditional training scheme, (C)
one hidden and (D) two hidden layers by cross-learning training scheme with fixed learning sets,
    (E) one hidden and (F) two hidden layers by cross-learning training scheme with randomly
                                     selected learning sets.

Results of ANN models based on different spectral data
The aforementioned three learning schemes of ANN were conducted with Savebest strategy to
record the best ANN model during the training process. The Savebest strategy is effective to
achieve the best results without over-fitting that frequently occurred in a traditional training
scheme. Therefore, the strategy was attempted in the present study to obtain the best model
from the oscillating convergence characterized by the proposed cross-learning scheme. Figure
4 shows a typical modeling results (SEC and SEV) of 5000 training epochs, the SEV of the
proposed model was converged at 8 mg/g.

            Figure 4. Distributions of SEC and SEV values in 5000 training epochs.
ANN models with the same initial values of weightings and thresholds will finally lead to a
constant error convergence at certain epochs (Table 4, the last column), but the modeling errors
with randomly selected learning sets did not converge at a defined epochs number. As shown
in the data rows for the proposed scheme, the standard deviations (as compared with the
averages) are in an acceptable range from four repeated modelings. Considering the error
convergences at every training epoch of the three ANN training schemes, the third type of ANN
with Savebest was suggested according to prediction ability. The SEV obtained with four input
nodes from spectra source of (400-2500, 2) was reduced from 10.20 mg/g (traditional ANN) to
8.84 mg/g (the proposed scheme). As compared with the SMLR result (12.35 mg/g), the
improvement is even more significant (28%).
The spectra source for calibration model is crucial for developing a practical multi-spectral
imaging system in agriculture applications. The ANN model with input data at 566, 574, 1396,
and 1530 nm was the best but requires expensive imaging devices as mentioned early. The
ANN results with four inputs from the spectra sources of shorter wavelengths, (450-950, 2) and
(450-950, 10), were more promising for a general multi-spectral imaging applications; the
accuracy were both improved 10 % of error convergence as compared with SMLR. The error
convergences obtained by the proposed ANN modeling with four inputs were not significantly

affected by their spectra sources. The choosing of spectral source for the ANN modeling is
therefore more flexible for constructing an economic and practical multi-spectral imaging system.
The proposed ANN models with seven inputs were also performed. Compared with the SMLR
results, the prediction accuracy had significant improvement except for the SEC from the
spectra source of (400-2500, 2). Additional input terms until seven terms did increase the
accuracy of SMLR analysis without over-fitting problems, but similar effects did not happen to
the proposed ANN model. Using non-linear transfer function to fit the error of nitrogen prediction
model, the ANN model with four selected input nodes obtained comparable results with seven
input nodes. The ANN model with four input nodes from spectra of (450-950 (10)) was
suggested for building the remote sensing model, the results, rc2=0.87, SEC=8.73 mg/g,
rv2=0.84, SEV=9.60 mg/g, were satisfactory.
Compared with SMLR, the proposed ANN is more effective in error-fitting ability, more flexible in
parameter (wavelength) selection. To monitor the nitrogen status of cabbage seedlings, multi-
spectral imaging system with silicon CCD camera and four band-pass filters of 490 nm, 570 nm,
600 nm, and 680 nm is suggested.
Table 4. Results of ANN with different training schemes.
 Spectra       No*      ANN        rc 2        SEC       rv2       SEV         Epochs
 source                 type
 (400-2500, 2) 4        ANNa       0.87        8.36      0.83      10.20       51
                        ANN        0.94        5.96      0.81      10.78       98
                        ANNc       0.89±0.014 8.27±0.19 0.86±0.003 8.84±0.15
 (450-950, 2) 4         ANNa       0.86        8.77      0.73      12.54       38
                        ANNb       0.88        8.28      0.74      12.30       103
                        ANNc       0.87±0.004 8.74±0.17 0.78±0.009 11.20±0.22
 (450-950, 10) 4        ANNa       0.91        6.86      0.80      10.82       100
                        ANNb       0.88        8.46      0.83      10.08       106
                        ANNc       0.87±0.010 8.73±0.23 0.84±0.006 9.60±0.23
 (400-2500, 2) 7        ANNa       0.86        8.68      0.79      10.91       14
                        ANNb       0.87        8.38      0.79      10.88       32
                        ANNc       0.88±0.010 8.25±0.34 0.85±0.014 9.39±0.40
 (450-950, 2) 7         ANNa       0.92        6.76      0.79      11.23       81
                        ANNb       0.89        7.85      0.78      11.10       53
                        ANNc       0.89±0.005 7.99±0.16 0.85±0.004 9.37±0.05
 (450-950, 10) 7        ANNa       0.86        9.68      0.79      10.91       14
                        ANN        0.87        8.38      0.79      10.88       32
                        ANNc       0.87±0.011 8.66±0.40 0.85±0.009 9.25±0.27
  Number of Wavelengths.
  Result of ANN with traditional learning scheme.
  Result of ANN with cross-learning scheme and fixed learning set.
  Average result of ANN with cross-learning scheme and random selecting sample for learning

Significant wavelengths for estimating nitrogen content of cabbage seedling leaves were
determined by SMLR analysis, and a modified ANN model was proposed to further increase the
prediction accuracy. Wavelengths were selected from several spectra sources including full
band spectra, spectra ranged from 450 nm to 950 nm, and spectra limited by commercially

available filters for multi-spectral imaging using CCD. The SMLR results using parameters of
four and seven wavelengths did not reach to linear regression equation with sufficient accuracy.
Using the strategies of cross-learning and random-sampling in/between the calibration sets, the
new training scheme was proven to be more precise than SMLR and traditional ANN methods.
The problematic over-fitting effect of ANN was extensively reduced, and data set with even
fewer parameters can be used to obtain satisfactory results from various spectra sources. A
multi-spectral imaging system for agricultural purposes can be easily constructed with CCD and
the band-pass filters when the proposed ANN model was adopted in the calibration modeling.

Bausch, W. C., and H. R. Duke. 1996. Remote sensing of plant nitrogen status in corn.
        Transactions of ASAE 39(5):1869-1875.
Boger, Z. 2003. Selection of quasi-optimal inputs in chemometrics modeling by artificial neural
        network analysis. Analytica Chimica Acta 490:31-40.
Boegh, E., H. Soegaard, N. Broge, C. B. Hasager, N. O. Jensen, K. Schelde, and A. Thomsen.
        2002. Airborne multispectral data for quantifying leaf area index, nitrogen concentration,
        and photosynthetic efficiency in agriculture. Remote Sens. Environ. 81:179-193.
Bolster, K. L., M. E. Martin, and J. D. Aber. 1996. Determination of carbon fraction and nitrogen
        concentration in tree foliage by near infrared: a comparison of statistical methods. Can.
        J. For. Res. 26:590-600.
Burks, T. F., S. A. Shearer, R. S. Gates and K. D. Donohue. 2000. Backpropagation neural
        network design and evaluation for classifying weed species using color image texture.
        Transactions of ASAE 43(4):1029-1037.
Diker, K., and W. C. Bausch. 2003. Radiometric filed measurements of Maize for estimating soil
        and plant nitrogen. Biosystem Engineering 86(4):411-420.
García-Ciudad, A., A. Ruano, F. Becerro, I. Zabalgoeazcoa, B. R. Vazquez de Aldana, and B.
        Garcia-Criado. 1999. Assessment of potential of NIR spectroscopy for the estimation of
        nitrogen content in grasses from semiarid grasslands. Animal Feed Science and
        Technology 77:91-98.
Gillon, D., C. Houssard, and R. Joffre. 1999. Using near-infrared reflectance spectroscopy to
        predict carbon, nitrogen and phosphorus content in heterogeneous plant material.
        Oecologia 118:173-182.
Goel, P. K., S. O. Prasher, J. A. Landry, R. M. Patel, A. A. Viau, and J. R. Miller. 2003.
        Estimation of crop biophysical parameters through airborne and field hyperspectral
        remote sensing. Transactions of the ASAE 46(4):1235-1246.
Goel, P. K., S. O. Prasher, R. M. Patel, D. L. Smith, A. DiTommaso. 2002. Use of airborne multi-
        spectral imagery for weed detection in field crops. Transactions of the ASAE 45(2):443-
Grossman, Y. L., S. L. Ustin, S. Jacquemoud, E. W. Sanderson, G. Schmuck, and J. Verdebout.
        1996. Critique of stepwise multiple linear regression for the extraction of leaf
        biochemistry information from leaf reflectance data. Remote Sens. Environ. 36:182-193.
Hsieh, C., Y. R. Chen, B. P. Dey and D. E. Chan. 2002. Separating septicemia and normal
        chicken livers by visible/near-infrared spectroscopy and back-propagation neural
        networks. Transactions of the ASAE 45(2):459-469.
Hsieh, K. W., S. Chen, W. H. Chang, M. T. Lee and C. T. Chen. 2001. A dynamic simulation
        model for seeding growth. Transactions of the ASAE 44(6):1949-1954.

Johnson, L. F., and C. R. Billow. 1996. Spectroscopic estimation of total nitrogen concentration
         in Douglas-fir Foliage. Int. J. Remote Sens. 17:489-500.
Kostrzewski, M., P. Waller, P. Guertin, J. Haberland, P. Colaizzi, E. Barnes, T. Thompson, T.
         Clarke, E. Riley, and C. Choi. 2003. Ground-based remote sensing of water and nitrogen
         stress. Transactions of the ASAE 46(1):29-38.
Liu, L., C. E. Goering, and L. Tian. 2001. A neural network for setting target corn yields.
         Transactions of the ASAE 44(3):705-713.
Mutanga, O., and A. K. Skidmore. 2004. Integrating imaging spectroscopy and neural networks
         to map grass quality in the Kruger Nation Park, South Africa. Remote Sensing of
         Environment 90(1):104-115.
Serrano, L., J. Penuelas and S. L. Ustin. 2002. Remote sensing of nitrogen and lignin in
         Mediterranean vegetation from AVIRIS data: Decomposing biochemical from structural
         signals. Remote Sens. Environ. 81:355-364.
Stone, M. L., J. B. Solie, W. R. Whitney, S. L. Taylor, and J. D. Ringer. 1996. Use spectral
         radiance for correcting in-season fertilizer nitrogen deficiencies in winter wheat.
         Transactions of the ASAE 39(5):1623-1631.
Ungar, S. G., J. S. Pearlman, J. A. Mendenhall, and D. Reuter. 2003. Overview of the earth
         observing one (EO-1) mission. IEEE Transactions on Geoscience and Remote Sensing,
White, J. D., C. M. Trotter, L. J. Brown, and N. Scott. 2000. Nitrogen concentration in New
         Zealand vegetation foliage derived from laboratory and field spectrometry. Int. J. Remote
         Sensing 21(12): 2525-2531.
Wood, G. A., J. P. Welsh, R. J. Godwin, J. C. Taylor, E. Earl, and S. M. Knight. 2003. Real-time
         measurement of canopy size as basis for spatially varying nitrogen applications to winter
         wheat sown at different seed rates. Biosystem Engineering 84(4):513-531.
Yoder, B. J. and R. E. Pettigrew-Crosby. 1995. Predicting nitrogen and chlorophyll content and
         concentrations from reflectance spectra(400-2500 nm) at leaf and canopy scales.
         Remote Sens. Environ. 53:199-211.


To top