c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/compag

Quantifying sub-pixel signature of paddy rice field using an

artificial neural network

Manoj Karkee, Brian L. Steward ∗ , Lie Tang, Samsuzana A. Aziz

Agricultural and Biosystems Engineering Department, Iowa State University, 100 Davidson Hall, Ames, IA, 50011, United States

a r t i c l e i n f o a b s t r a c t

Article history: Remote sensing (RS) can be used for regional and global scale monitoring of crop devel-
Received 24 July 2007 opment, crop health, and cropping practices and also for water resources planning and
Received in revised form designing. Lower spatial resolution, such as 1 km MODIS imagery, for example, is useful for
18 July 2008 national, regional and global scale studies, but sub-pixel mixing of different landuses may
Accepted 24 July 2008 occur. In the case of rice production, one pixel of such resolution may include land area with
rice grown under different cropping practices such as one, two and three crops per year rice.
Keywords: A method was developed for quantifying sub-pixel landuses of individual rice types using
Mixed pixel an artificial neural network (ANN). Temporal patterns of normalized differential vegetation
Artificial neural network index (NDVI) depend on and result from the complex relationship between NDVI and crop-
Remote sensing ping practice parameters associated with a pixel. In the case of a rice field, these parameters
Temporal vegetation signature consist of the area fractions of different types of rice and their emergence dates. An ANN
Rice cropping practice was used as a model inverter to estimate these parameters. The data for this research were
produced numerically using the soil–water–atmosphere–plant (SWAP) crop growth model.
Crop area fractions within a pixel were predicted with an RMSE of 1.3% and an average
estimated emergence date error of 4 days. A representative experiment conducted using a
2.4 GHz Pentium processor based desktop computer took about 1.22 ␮s per pixel, which was
substantially faster than the genetic algorithm based approach of Ines and Honda [Ines,
A.V.M., Honda, K., 2005. On quantifying agricultural and water management practices from
low spatial resolution RS data using genetic algorithms: a numerical study for mixed-pixel
environment. Advances in Water Resources 28 (8), 856–870]. The ANN based approach was
computationally very efficient and thus practical to apply to satellite imagery consisting of
millions of pixels.
© 2008 Elsevier B.V. All rights reserved.

1. Introduction water use (Ines, 2002) because irrigated agriculture consumes

almost 70% of the total water used in the world (Seckler et
Increasing global population results in increasing water al., 1999). Optimal design of irrigation networks and reservoirs
demand for irrigation, domestic and industrial uses. Sustain- helps promote sustainable water use by increasing irrigation
able use of water among multiple users is essential to meet water use efficiency (Paudyal et al., 1991).
the ever-increasing demand with limited resources. Irrigation Researchers have been working on optimizing the size
is one of the biggest potential areas for promoting sustainable and layout of reservoirs and canal networks (Tod et al., 1990;

∗
Corresponding author at: Agricultural and Biosystems Engineering Department, Iowa State University, 214D Davidson Hall, Ames, IA,
50011, United States. Tel.: +1 515 294 1452; fax: +1 515 294 2255.
E-mail address: [email protected] (B.L. Steward).
0168-1699/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.compag.2008.07.009
66 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

Paudyal et al., 1991; Lejano, 2006) while meeting the rate, (Tateishi et al., 2004). Several non-linear mixing model based
duration and frequency of water delivery required by the techniques have also been developed (Lu et al., 2003; Foody et
farmer. However, the effectiveness of these optimization tech- al., 1997), which generally outperform the linear model based
niques depends on how accurately the spatial and temporal studies but at the cost of increased computation (Liu and Wu,
water demands are estimated (Tod et al., 1990). In south 2005). Several other unmixing techniques have been proposed,
and south-east Asia where rice is the predominant irriga- which are based on the soft computing algorithms such as
tion water user, information about rice cropping practices fuzzy logic (Foody and Cox, 1994), genetic algorithms (GA) (Ines
essential for accurate estimation of water demands is either and Honda, 2005) and artificial neural networks (ANNs) (Liu et
unavailable or inexact (Malingreau, 1986a). Remote sensing al., 2004; Foody et al., 1997).
(RS) based studies of crop types and cropping practices have The focus of previous sub-pixel unmixing studies has been
the potential to help irrigation water planners and design on the classification of pixels into fractions of different lan-
engineers accurately estimating those demands (Xu et al., duse/landcover classes (Shimabukuro and Smith, 1991). In
2006). agricultural monitoring, the mixed pixel problem consists
Over the years, remote sensing has been investigated as of both landuse fraction estimation and the estimation of
a means of monitoring agricultural and water management temporal parameters such as sowing or emergence dates
activities because both the spatial and temporal characteris- associated with different cropping practices. Even though
tics of a region can be easily obtained by RS imagery (Droogers the classification principle remains the same, including tem-
and Bastiaanssen, 2002; Hunsaker et al., 2005). As a result, sea- poral parameters estimation into the problem increases its
sonal changes in vegetative growth at the regional level can be complexity. Ines and Honda (2005) showed that it is possi-
monitored to help policy makers and farm/water managers ble to estimate landuse fraction and emergence date from
make better decisions (Ines and Honda, 2005). In most cases, the temporal patterns of low spatial resolution data. Using
crop monitoring using RS has been practiced with high resolu- the soil–water–atmosphere–plant (SWAP) model to simulate
tion remote sensing data. However, high resolution techniques temporal vegetation signatures, they formulated the unmix-
are beneficial only to small (field) scale applications and have ing task as an optimization problem and successfully applied
higher cost and lower temporal resolution. Low spatial res- a genetic algorithm to estimate these rice field parameters
olution data offers higher temporal resolution, wider spatial from temporal evapotranspiration (ET) and leaf area index
coverage, and minimal cost as compared to the high spa- (LAI) patterns. The technique was limited, however, due to the
tial resolution data, which makes it more applicable to larger computational time it took, 1.5 h, for the GA with a population
(regional and global) scale agricultural monitoring (Takeuchi size of five individuals to evolve the temporal pattern of ET or
et al., 2003). LAI of a single pixel. It is difficult to apply this technique, in
Low resolution remote sensing data products are avail- its given form, to a practical image, which typically consists of
able publicly. For instance, several Terra Moderate Resolution millions of pixels.
Imaging Spectroradiometer (MODIS) products are available Since sub-pixel estimation of cropping practices could be
free of cost through the USGS and other public organiza- viewed as a soft modeling problem to find the relationship
tions. USGS provides daily global imagery at 250 m, 500 m, between temporal vegetation signatures and cropping prac-
and 1000 m spatial resolutions. Eight-day composite MODIS tices, an artificial neural network could be used. An ANN
Surface Reflectance products (MOD09A1) are also available. is a generic modeling tool, which captures and models the
MODIS composite data is generated by first eliminating pix- underlying relationship between input and output variables
els with a low score or low observational coverage and then represented by a number of training input/output pairs. Neu-
choosing an observation with the minimum blue-band value ral networks learn new knowledge by adjusting connection
during the 8-day period (Xiao et al., 2006). A composite normal- weights between nodes to produce a specified output. New
ized differential vegetation index (NDVI), which is one of the outputs will be consistent with previous experience stored
commonly used indices for agricultural monitoring, can then in the connection weights (Diane et al., 1995). ANNs have
be derived from the composite reflectance imagery (Rouse et widely been applied in remote sensing image interpretations.
al., 1973). Typically, there are two categories of applications: classifi-
One major challenge with the low resolution data, how- cation (Diane et al., 1995; Chen et al., 1997; Foody et al.,
ever, is the problem of land cover mixing inside a pixel. It is 1997; Ito and Omatsu, 1997; Brivio et al., 2003; Liu et al.,
likely that a pixel contain several agricultural landuse types 2004) and modeling and/or inverse modeling (Chen et al.,
because of the low spatial resolution (Ines and Honda, 2005). 1993; Delfrate and Wang, 2001). ANNs have also been applied
Sub-pixel level estimation of different landuse types with spa- more generally in agriculture. Applications vary from irriga-
tial resolution of less than the resolution of the remote sensing tion management (McClendon et al., 1996) to classification of
data is necessary for accurate landcover mapping with remote weeds, fruits and vegetables in machine vision (Tang et al.,
sensing data (Liu and Wu, 2005). 2003; Shahin et al., 2002) to precision farming (Irmak et al.,
Sub-pixel classification methods have been developed 2006). When it comes to sub-pixel unmixing, ANNs are not
based on both linear and non-linear mixing models. How- only superior to the traditional linear unmixing algorithms,
ever, linear mixing models have been used the most (Settle but also one of the superior non-linear classifiers (Liu and Wu,
and Drake, 1993; Bosdogianni et al., 1997; Roberts et al., 1998; 2005).
Tateishi et al., 2004). In these models, it is assumed that the The inputs to the system proposed by Ines and Honda
pixel value is a linear combination of sub-pixel component (2005) were ET and LAI. However, these biophysical param-
reflectances with coefficients proportional to component area eters are estimated using models relating these parameters
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76 67

Fig. 1 – Flow chart of the research. Upper half part of the flowchart shows the procedure to generate numerical data and the
lower half part shows the use of the ANN to invert the relationship between the NDVI pattern and cropping parameters of a
pixel.

to remote sensing data whereas vegetative indices like NDVI

are calculated directly from the data using a mathematical
2. Materials and methods
combination of spectral bands. Seasonal pattern of compos-
Temporal composite NDVI patterns were used to estimate dif-
ite NDVIs have been more widely used in both full-pixel
ferent agricultural practices within a pixel of low resolution
as well as sub-pixel landuse/landcover classification. Sev-
remote sensing data. The method of estimating sub-pixel agri-
eral full-pixel classification studies have been carried out to
cultural practices started with the generation of the training
classify landuse based on 1-year-long patterns of composite
and testing datasets (Fig. 1) using the SWAP model and a rela-
NDVI (Defries and Townshend, 1994; Sugita and Yasuoka, 1996;
tionship between LAI and NDVI. A temporal NDVI pattern and
Keiko et al., 1999; Lacoul et al., 2000; Xiao et al., 2006). Simi-
the corresponding cropping parameters were used to define an
larly, many sub-pixel unmixing studies have also used NDVI
input–output pair. These input–output pairs were used to train
patterns (Tateishi et al., 2004; Fisher, 1994; Uchida, 2001). These
an ANN to model the input–output relationship described
studies show that the composite temporal NDVI patterns
by the training dataset. Finally, the performance achieved by
could be the better choice in estimating sub-pixel cropping
the ANN was assessed using RMSE and relationship between
practices.
actual and predicted values of the output parameters.
The objectives of this study were to estimate landuse
fractions of different cropping practices and crop emergence
dates within single pixels of low spatial resolution remote 2.1. The SWAP model
sensing data using an ANN and temporal patterns of com-
posite NDVIs, and to compare the performance of the ANN SWAP is a commonly used computer model, which is based
approach with the GA-based approach of Ines and Honda on the direct interaction between water flow, solute transport,
(2005). heat flow, and plant growth to simulate physical, chemical and
68 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

the premises of Asian Institute of Technology, Klong Luang,

Pathumthani, Thailand. The data were collected daily for a
year.

2.2. Generating crop parameters

Temporal composite NDVI patterns of a pixel were used to esti-

mate different cropping practices within a pixel. As in Ines and
Honda (2005), a typical rice cropping system was chosen for
further investigation. Three types of cropping practices were
simulated: one crop per year rainfed rice represented by the
label, RICE1; two crops per year irrigated rice represented by
Fig. 2 – Daily temperature profiles of a year used in the RICE2; and three crops per year irrigated rice represented by
SWAP model. Data was collected by Ines and Honda (2005). RICE3. These three are the standard cropping practices in the
paddy rice fields of south and south-east Asia (Malingreau,
1986b; Ines and Honda, 2005; Xiao et al., 2006; Pons, 2008) and
biological processes (SWAP, 2008). The model is based on the they are planted in standing water. There could be some other
one dimensional Richards’ equation, which uses the spatial landuses, such as small patches of grass and other vegetation,
and temporal potential differences in the soil column to cal- small built-up areas and jungles within some of the pixels of
culate the soil water movement (Van Dam et al., 1997; Ines paddy field. These landuse patches were assumed to be negli-
and Honda, 2005). The model has been validated and widely gible in this study. In such an environment, the NDVI pattern of
used over the past two decades (Abenney-Mickson et al., 1997; a pixel observed from remote sensing depends upon the frac-
Tedeschi and Menenti, 2002; Ines and Honda, 2005). tions of the landuses of three different types of rice, sowing
The inputs to the SWAP model were emergence and harvest or emergence dates of each type of rice, soil parameters, and
dates of a crop, weather data, and soil parameters. The output weather parameters (Fig. 4). Thus, a pattern of crop param-
is the temporal pattern of daily LAI and ET values. The weather eters consisted of three area fractions of a pixel covered by
data (Figs. 2 and 3) and soil parameters (Table 1) were same RICE1, RICE2 and RICE3 (represented by a1 , a2 and a3 ), emer-
as Ines and Honda (2005) and made available to us through gence date of RICE1 (represented by e11 ), two emergence dates
the Honda Lab, Asian Institute of Technology, Thailand. The of RICE2 (represented by e21 and e22 ), and three emergence
weather data were collected at the weather station located in dates of RICE3 (represented by e31 , e32 and e33 ).

Fig. 3 – Daily profiles of rain fall, wind velocity, humidity and solar radiation of a year used in the SWAP model. Data was
collected by Ines and Honda (2005).

Table 1 – Input soil parameters to the SWAP model (Ines and Honda, 2005)
Soil parameters Layer 1 (0–60 cm) Layer 2 (60–200 cm)
−1
Shape parameter of drying and wetting curve, ˛ (cm ) 0.021 0.018
Shape parameter of the hydraulic conductivity function,  −2.523 −3.997
Residual moisture content,  res (cm3 /cm3 ) 0.000 0.000
Saturated moisture content,  sat (cm3 /cm3 ) 0.384 0.365
Saturated hydraulic conductivity, Ksat (cm/day) 15.282 10.188
Shape parameter of soil water retention curve, n 1.208 1.137
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76 69

Fig. 4 – Data flow schematic of the research. The box in the left represents the simulated NDVI patterns for three classes of
rice. The pattern in the right is the mixed pixel signature.

The emergence dates of RICE1 and RICE2 vary significantly from a uniform distribution and ranging from mid January
over the fields depending upon the factors such as temper- to late February for the first emergence date and from late
ature and/or rainfall pattern. This variability in emergence May to early July for the second emergence date subject to the
dates is possible because the time gap available between two constraint:
consecutive crops is substantial. However, variability in the
emergence dates in a RICE3 field is very small because the time e22 ≥ e21 + 100 (days) (2)
gap available between two consecutive crops is very small and
the cropping does not depend on the rainfall pattern. Studies
This constraint eliminated crop overlap (Ines and Honda, 2005)
have reported that in the fields where three crops are planted,
because the rice crop takes at least 100 days from the emer-
one crop follows the other immediately. Generally the first
gence date to be ready for harvesting.
crop of a growing season is planted in early May, the second
in late August or early September and the third in early to
mid January (Malingreau, 1986b; Ines and Honda, 2005; Pons, 2.3. Simulated sub-pixel mixing data
2008). So, to simplify the problem, the emergence dates for
RICE3 were fixed on the 1st of January, 1st of May and 1st of The SWAP model simulating remote sensing data was used
September. This simplification reduced the complexity of the to study the performance of the method before committing to
network and the computational time required in both training the resources for field data collection. A study using simulated
and testing phases. data required not only reduced time and cost, but also pro-
One hundred patterns of the crop fractions were randomly vided an opportunity to study and evaluate the performance
generated from a uniform distribution subjected to the con- of the technique in various extreme conditions. One hundred
straint: cropping pattern emergence dates for RICE1, RICE2 and RICE3
were provided to the SWAP model along with weather data
and soil parameters to generate daily values of LAI over a year.
a1 + a2 + a3 = 1.0 (1)
Generally, low spatial resolution remote sensing data are pro-
vided in 7–30-day composites. To represent the simulated data
which means the area covered by individual rice cropping in the form similar to the real remote sensing data, daily LAI
practices should sum up to the total area. The emergence date values were averaged over the 10 days intervals to generate the
for RICE1 was allowed to vary from mid April to late June. Simi- temporal patterns of 10-day composite LAI. Then, the LAI com-
larly, the emergence dates for RICE2 were randomly generated posites were mixed according to the area fractions a1 , a2 and
70 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

corresponding to patterns, and each of the columns sequen-

tially representing one of the composite NDVI values making
up the temporal patterns. The corresponding inputs of emer-
gence dates and crop area fractions and output NDVI patterns
were then integrated to define the input–output dataset for
the neural network.
To test model robustness, further experiments were car-
ried out by adding white noise to the simulated NDVI patterns.
White Gaussian noise with a zero mean and 5% standard devi-
ation was added to the NDVI values.

2.5. Inversion of the temporal NDVI profiles using

ANN

Fig. 5 – Sample LAI and NDVI patterns generated 2.5.1. Preprocessing

numerically using SWAP model and a relationship between Two preprocessing techniques were applied to the input
LAI and NDVI given by Xiao et al. (2002). dataset: the standard normal variate (SNV) transformation
(Barnes et al., 1989) and principle component analysis (PCA).
The SNV transformed the dataset into zero mean and unity
a3 and the linear mixing model described in Ines and Honda standard deviation patterns. PCA reduced the dimensionality
(2005). of the dataset from 36 to 8 while preserving 99% variability of
the dataset. Pattern preprocessing leads to an equal learning
2.4. Generating temporal NDVI patterns rate and minimizes weight synchronization during network
training (Haykin, 1994).
Each 1-year long output LAI pattern (Fig. 5) was represented
by a sequence of 36 LAI values, each value representing an 2.5.2. Neural network design
average of 10 daily simulated values. Since the input to the A feedforward multilayer perceptron with error backpropaga-
ANN was an NDVI pattern, the LAI patterns were transformed tion neural net (MLPNN) was used as the model inverter. The
into NDVI patterns using the relationship first patterns applied to the network were the 36-element long
√ SNV transformed temporal NDVI patterns. Second, the eight
NDVI = 0.065LAI + 0.153 LAI + 0.09 (3) dimensional PC patterns were applied to the network. Con-
sequently, two different MLPNN architectures were developed
This relationship was particularly suitable for this study as it with 36 and 8 input nodes to accommodate the two patterns.
was developed for the rice field using 10-day composite 1KM The output layer of each network consisted of six nodes to map
spatial resolution SPOT 4 satellite data (Xiao et al., 2002). The an output pattern consisting of the three emergence dates and
NDVI patterns derived from the LAI patterns closely followed the three area fractions. Target outputs were transformed into
the emergence/growth cycle of the cropping practice observed the +0.15 to +0.85 unipolar range for the emergence dates and
from the simulated LAI patterns (Fig. 5). the 0–1 range for the crop fractions. With this transformation,
Remote sensing data may be affected by sun angle and sen- a value of 0.15 represented an emergence date of the first day
sor viewing angle at the time of imaging. Also, the weather and of a year and a value of 0.85 represented an emergence date
ground condition such as cloud cover, standing water in the of the last day (365th day) of a year. Anything inside the range
rice fields, and soil moisture may affect reflectance. Because of 0.15–0.85 was linearly related to the emergence dates; any-
Xiao et al. (2002) did not explicitly report about the sun angle thing outside the range saturated to the beginning and ending
at the time of image acquisition it was not possible to include of the year end-points. Reduction of the unipolar range was
the sun angle effect in the study. We assumed the relationship necessary to cope with the dynamic range of sigmoid activa-
was developed from the data adjusted for the bidirectional tion function, because the response of this activation function
reflectance distribution function (BRDF) effects. Nadir-BRDF saturated at the limiting values.
adjusted reflectance would be the best data to be used in The 100 simulated NDVI patterns were divided into three
the application of this technique with real field data. USGS sets: 50 were used as the training patterns, 25 as the valida-
has already begun producing such products (e.g. nadir-BRDF tion patterns, and the remaining 25 as the testing patterns.
adjusted MODIS reflectance). Users would not need any fur- The training patterns were used in the training process to
ther preprocessing for BRDF, if such products are used. Cloud adjust the ANN connection weights. Validation patterns were
cover effects were assumed to be minimal because the data not applied for backpropagation training but were used to stop
were 10-day composite NDVI patterns. Soil moisture and soil the training process. RMSE was calculated over these patterns
reflectance were also assumed not to have a substantial effect in each epoch, where an epoch is defined as the presenta-
on the NDVI as rice is planted in standing water, and the field tion of the complete set of training patterns to the neural
is almost always covered by the water (Xiao et al., 2002). network. Training was stopped when the accuracy over the
The NDVI dataset consisted of 100 NDVI patterns, each pat- validation patterns became saturated or started degrading.
tern represented by 36 data points taken over a year at a 10-day This stopping criterion helped avoid network overfitting the
interval. The patterns were stored in an array with the rows training data. Testing patterns were used to independently
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76 71

Table 2 – Comparison of network performance with different architectures

Hidden layers Dataseta Network Training Training RMSEc training RMSE RMSE
architectureb epochs time (ms) validation testing

1 PCA 08:04:06 165 64 0.033 0.043 0.042

08:07:06 139 78 0.035 0.049 0.045
08:14:06 160 201 0.036 0.053 0.061
Full 36:05:06 95 98 0.030 0.051 0.051
36:10:06 65 140 0.026 0.047 0.049
36:21:06 159 685 0.025 0.053 0.056

2 PCA 08:04:03:06 199 97 0.032 0.043 0.045

08:07:07:06 139 136 0.033 0.050 0.049
08:14:12:06 160 350 0.030 0.047 0.048
Full 36:05:04:06 180 229 0.024 0.046 0.048
36:15:10:06 93 351 0.024 0.044 0.043
36:26:16:06 148 1040 0.021 0.043 0.039

a
Full length pattern has 36 NDVI values and the PCA pattern has 8 NDVI values per pattern.
b
Network Architecture representation is number of input nodes:hidden layer 1 nodes:[optional hidden layer 2 nodes]:output nodes.
c
RMSEs were calculated based on the normalized data.

assess the accuracy of the trained model. The training, val- fixed at the best value and then the momentum term was var-
idation and testing sets were randomly selected, and the ied from 0.0 to 1.0, and network performance was determined
training examples were presented to the network in a ran- in each experiment. In these experiments, the speed of con-
dom order, which is critical for improving convergence speed vergence was determined by the number of training epochs
(Haykin, 1994). taken by the network.
The RMSE was calculated based on the normalized data
using the common RMSE formulae from statistics, which takes
2.5.3. Experiments on different architectures and training
the square root of the mean of the squared errors of the out-
parameters
puts. Since two different types of outputs (area fractions and
Two sets of experiments were carried out to investigate
emergence date fractions) were included in the RMSE calcu-
the effect of network architecture and training parame-
lation, the result did not have any physical unit. However,
ters on network performance. First, network architecture
RMSE could easily be interpreted in terms of percentage of
was varied by the number of hidden layers (1 and 2),
the output parameters, percentage of the total area fraction
number of input neurons (8 and 36), number of first hid-
and percentage of a year in the case of emergence dates.
den layer neurons (4, 5, 7, 10, 14, 21 and 26) and the
Finally, the network with the best architecture, learning
number of second hidden layer neurons (3, 4, 7, 10, 12
rate, and the momentum found in the prior experiments was
and 16). In total, 12 such combinations were investigated
used to compare the performance of logistic sigmoid and hyper-
(Table 2). The training, validation and testing RMSEs were
bolic sigmoid activation functions, and Levenberg–Marquardt
recorded for each experiment. The number of epochs and
backpropagation and gradient decent with momentum backpropa-
training time were also recorded. The architectures were
gation training methods. The combination of best performing
compared based on the computational time taken and
activation function and training method was selected for fur-
the validation and testing accuracies achieved by the net-
ther use in this research.
works. Network architectures were represented by input
nodes:hidden layer 1 nodes:[optional hidden layer 2 Nodes]:output
nodes. 3. Results and discussion
The best performing network architecture was used in the
second set of experiments on the effects of varying learning As described in the methodology, the effect of one or two
rates and momentum terms. Learning rate can be thought of hidden layers and the number of neurons in each hidden
as the acceleration and momentum as the speed of the learn- layer on the network performance was assessed (Table 2). As
ing process. Smaller learning rates will reduce the changes there were more inputs in the full length patterns than the
to the synaptic weights in the network from one epoch to PCA shortened patterns, the ANN needed many more con-
the next and will smooth the trajectory in weight space. The nections within the network to model the relationship with
momentum term in the delta rule helps avoid oscillation even the full length patterns. Two hidden layers with a large num-
with the larger value of learning rate (Haykin, 1994) and helps ber of nodes in each layer were necessary to achieve a good
track the global minimum in case the training process could performance. It was found in this case that the accuracy of
be trapped in local minima. the network kept decreasing with the decreasing numbers
First, the learning rate was varied from 0.0 to 1.0 in 0.1 of hidden layers and the number of nodes in each hidden
intervals, and network performance was determined in each layer. All the experiments with two hidden layers had lower
experiment by evaluating the speed of convergence and the RMSEs than with one hidden layer. In general, the valida-
training and validation RMSEs. The momentum term was tion and testing RMSEs decreased with increasing number of
fixed at 0.0 in these experiments. Next, the learning rate was hidden layer neurons. In contrast, one hidden layer with rel-
72 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

Fig. 6 – Effect of learning rate on network performance. As

Fig. 7 – Learning curves for varying momentum term. For
the learning rate increases, convergence speed of the
the range of 50–200 numbers of epochs, momentum of 0.0,
network also increases (generally decreasing number of
0.1 and 0.5 achieved similar accuracy.
epochs) at the cost of reduced error performance
(increasing training RMSE).

rate throughout the entire range. Higher accuracy was con-

atively fewer neurons was found to be adequate for the PCA sidered more important than training time, since the training
reduced patterns. In this case, increasing the number of hid- session needed to be run only once for the many feedforward
den layers and the number of nodes in each hidden layer did runs that followed. The learning rate selected was 0.1, slightly
not improve accuracy, but rather decreased accuracy in some smaller than the value of 0.2 at the minimum of the epoch
cases. This result indicated that PCA effectively reduced the curve.
dimensionality of the input data and thus the complexity of In the range of 50–200 epochs, the network was insensitive
the network without loss of performance. In the pool of net- to the momentum values in the range of 0.0–0.5 (Fig. 7). The
works with both full length patterns and PCA shorten patterns, network performance was the same for all the momentum
the two network architectures that achieved the lowest error values in this range. However, at the momentum of 0.9, the
were the 08:04:06 network and the 36:26:16:06 network. The network performance degraded substantially. Momentum of
validation and testing RMSEs achieved by the former archi- 0.1 was chosen for further study.
tecture was 0.043 and 0.039, and that achieved by the later When it comes to activation functions, the logistic sig-
was 0.043 and 0.042, respectively, which were not substantially moid function performed better than the hyperbolic sigmoid
different. However, the 08:04:06 network was much simpler because the unipolar response of logistic sigmoid better
than the 36:26:16:06 network. Consequently, the former archi- matched the output format. The output layer activation func-
tecture took only 64 ms where as the latter took 1040 ms to tion was pure linear, which prevented the restricting of the
converge during training. So, the 08:04:06 network architecture output range by a sigmoid function. In case of training meth-
was found to have the best performance in terms of accuracy ods, both Levenberg–Marquardt backpropagation and gradient
achieved and training time and used in the further experi- decent with momentum backpropagation achieved similar
ments. performance. Due to its simplicity and robustness, gradient
Larger learning rates generally led to faster convergence at decent with momentum was selected for the further study.
the cost of degraded performance (Fig. 6). Starting at a learn- The combination of best performing network architec-
ing rate of 0.0, the number of epochs needed to converge ture (08:04:06) and training parameters (learning rate 0.1,
decreased with increasing learning rate until a minimum momentum 0.1, logistic sigmoid activation function, linear
of 35 was reached at a learning rate of 0.2. As the learn- output layer function and gradient decent with momentum
ing rate increased, the number of epochs increased slowly backpropagation training method) was used to evaluate and
before stabilizing at a value of about 55. However, training summarize individual emergence date and area fraction errors
RMSE monotonically increased with the increasing learning (Table 3). The ANN inverse modeling technique accurately esti-

Table 3 – Test errors produced by the best network configuration for the six output variables
Error measures Emergence dates (number of days) Crop area fractions (%)

RICE1 1st, RICE2 2nd, RICE2 RICE1 RICE2 RICE3

RMSEa 14 4 10 1.3 1.5 1.4

Absolute maximum error 27 10 21 3.5 4.7 3.3

a
RMSEs were calculated based on the normalized data.
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76 73

measure, the RMSE, for the same was 14 days. The maxi-
mum error for the first emergence date of RICE2 was only
10 days and that of the second emergence date of RICE2 was
21 days. The emergence date errors of RICE1 were found to
be bigger than those of RICE2. It was probably because of
the wider variation of the RICE1 emergence date in training
dataset. The maximum errors in the estimation of the lan-
duse fraction of RICE1, RICE2, and RICE3 were 3.5%, 4.7%, and
3.3%, respectively. Though the results were good, whether
or not they are acceptable will depend upon the type of the
application.
In another set of experiments, the same neural network
architecture and training parameters were applied to a dataset
with added white noise. The network converged and produced
about the same training error as observed with clean data.
As expected, the noise slightly affected the accuracy of the
testing results (Table 4). The prediction error was always less
than 4.6%. The network predicted the emergence dates with
an error of 17 days, 8 days and 11 days for RICE1, the first
emergence date of RICE2 and the second emergence date of
RICE2, respectively. Similarly, the RMSEs were 3.0%, 2.7% and
1.6%, respectively, for the area fractions of RICE1, RICE2 and
RICE3.
To compare the results with the GA-based mixed pixel
results reported by Ines and Honda (2005), mean and standard
deviation of different cropping parameters of a pixel were cal-
culated (Table 5). With the clean dataset, the network achieved
the mean errors of 0 days, 0.4 days and 1 day with the stan-
dard deviations of 16.1 days, 3.5 days and 9.7 days, respectively,
for the emergence date of RICE1, and the first and the second
emergence dates of RICE2. The mean errors in the estimation
of the landuse fraction of RICE1, RICE2, and RICE3 were all
zero with standard deviation of 1%, 2% and 1%, respectively.
With the noisy dataset, the mean errors achieved were 7.7
days, 0.7 days and 0.4 days for the emergence date of RICE1,
and the first and the second emergence dates of RICE2 with
the standard deviations of 24.2 days, 7.3 days and 11.3 days,
respectively. Similarly, the mean errors of the landuse frac-
tions were 0%, 1% and 2%, respectively, for RICE1, RICE2 and
RICE3 with standard deviations of 6% each. The performance
of the ANN was slightly better than that of the GA with the
clean dataset. For the noisy dataset, the ANN achieved slightly
lower errors for crop area fractions and slightly higher errors
Fig. 8 – Relationship between actual and predicted crop
for the emergence dates.
area fraction of, (a) RICE1, (b) RICE2, and (c) RICE3.
Computationally, the neural network based approach was
much less costly relative to the GA-based approach suggested
mated the crop area fractions and emergence dates of three by Ines and Honda (2005). To estimate the parameters of one
rice cropping practices in a mixed pixel environment. The net- pixel, this approach took 1.22 ␮s in a candidate experiment
work predicted the emergence date of RICE1 with an RMSE of conducted in a desktop computer with a Pentium 4 micro-
14 days, the first emergence date of RICE2 with 4 days, and the processor running at 2.4 GHz with 512 MB RAM. This was a
second emergence date of RICE2 with 10 days. As mentioned tremendous improvement over the time taken by GA-based
before, the emergence dates for RICE3 were considered to be approach, which took about 1.5 h to complete 150 generations
fixed. The network achieved a RMSE of 1.3% of the pixel area in of a population of five in a Pentium 4 processor, 1.8 GHz speed
predicting the crop area fraction for RICE1, 1.5% for RICE2, and with 256 MB RAM (Ines and Honda, 2005). Though the hard-
1.4% for RICE3. The predicted area fractions plotted against the ware used in this study was faster by a factor of about two,
actual area fractions for all three rice showed a very good fit the computational speed achieved was 4.4 × 109 times faster
(Fig. 8). The r2 values for the actual versus predicted crop area than the GA-based study. The training time of the candidate
fractions of RICE1, RICE2, and RICE3 were all equal to 0.99. network was 64 ms. This improvement in the computational
The maximum error on estimating RICE1 emergence date speed opens up the opportunity for practical applications of
was almost 4 weeks. However, the more meaningful error the method.
74 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

Table 4 – Test errors produced by the network for the dataset with added white noise
Error measures Emergence dates (number of days) Crop area fractions (%)

RICE1 1st, RICE2 2nd, RICE2 RICE1 RICE2 RICE3

a
RMSE 17 08 11 3.0 2.7 1.6
Absolute maximum error 38 14 29 6.1 5.3 3.3

a
RMSEs were calculated based on the normalized data.

Table 5 – Comparison of solutions of ANN with genetic algorithm based study of Ines and Honda (2005)
Clean data Noisy data

Mean error S.D. Mean error S.D.

GA ANN GA ANN GA ANN GA ANN

Emergence date, RICE1 (day) 21 0 N/A 16.1 4.4 7.7 8.78 24.4
First emergence date, RICE2 (day) 1 0.4 N/A 3.5 0.2 0.7 0.45 7.3
Second emergence date, RICE2 (day) 0 1 N/A 9.7 −0.6 0.4 2.7 11.3

Crop area fraction, RICE1 0.03 0 N/A 0.01 −0.03 0 0.05 0.06
Crop area fraction, RICE2 0.02 0 N/A 0.02 0.01 0.01 0.04 0.06
Crop area fraction, RICE3 0.01 0 N/A 0.01 0.02 0.02 0.01 0.06

From the comparison, it can be expected that the ANN emergence dates from a low resolution mixed pixel composite
approach would perform better than the GA approach both NDVI patterns. A number of tests were carried out with differ-
in terms of accuracy and computational speed. The ANN ent neural network architectures, learning rates, momentum
approach used NDVI patterns instead of LAI patterns used terms, activation functions, and training methods to inves-
by the GA approach. Because remote sensing based LAI is a tigate their effect on ANN performance. The performance of
derived index whereas NDVI is a directly calculated index, use the neural network approach was compared with the Ines and
of NDVI data could be more practical. One drawback with the Honda’s (2005) GA approach both in terms of mean and stan-
ANN approach, however, is that when real field data is used, dard deviation of the error and computation time required.
this approach requires training data in addition to the testing The following conclusions can be drawn from this research:
dataset needed by the GA approach. Additional data collection
would require extended field work and additional cost. • An ANN can estimate sub-pixel rice cropping practice
The ANN approach extracted temporal and spatial infor- parameters. The RMSE achieved for the rice area fraction
mation about rice crops that could be used to estimate estimation was always less than 1.5% with clean data and
temporal and spatial water demand due to crops and crop- was less than 2.9% for the dataset with added white noise.
ping activities. This knowledge could be used to predict future The prediction errors in emergence dates varied from 4 days
temporal and spatial distributions of water demand. These to 14 days without noise and 8 days to 17 days with added
estimates are essential for the design and sizing of new irriga- white noise. In majority of the cases, mean and standard
tion reservoirs and canal networks and to improve the water deviation of the error achieved by this method were smaller
distribution from the existing irrigation projects. than those achieved by the GA-based study.
Accuracy of the regional and global scale agricultural lan- • The ANN approach can perform better than the GA
duse mapping could be affected by the spatial and temporal approach both in terms of accuracy and computational
resolution of the satellite dataset. Low resolution satel- speed. With the clean dataset, the mean errors achieved by
lite images like MODIS and Advanced Very High Resolution the ANN were less than 1 day for the emergence dates with
Radiometer (AVHRR) are the primary source of data used the standard deviations within 16 days. The mean errors
for regional and/or global scale landuse mapping. It is pos- for the landuse fraction were zero with standard deviations
sible that different landuse types will occur within a pixel within 2%. This performance of the ANN was slightly better
of such low resolution images. Mapping different landuse than that of the GA. Using a desktop computer, the ANN took
types at sub-pixel level could therefore be of great inter- 1.22 ␮s to estimate the sub-pixel parameters. This was a
est (Liu and Wu, 2005). The proposed sub-pixel cropping tremendous improvement over the time taken by GA, which
parameter estimation technique can be used to improve the was about 1.5 h in slightly older hardware configuration.
accuracy of regional and global scale agricultural landuse and The computational demands of the ANN approach were low
landcover maps. enough to be applicable to real images having millions of
pixels.

4. Conclusions Acknowledgements

A neural network based technique was developed to esti- This research of the Iowa Agriculture and Home Economics
mate the proportions of three rice cropping systems and their Experiment Station, Ames, Iowa, Project no. 3612, was sup-
 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76 75

ported by Hatch Act and State of Iowa funds. Executable mixed-pixel environment. Advances in Water Resources 28
programs, parameter files and related information for the (8), 856–870.
SWAP model were provided by Dr. Kiyoshi Honda, Asian Insti- Irmak, A., Jones, J.W., Batchelor, W.D., Irmak, S., Boote, K.J., Paz,
J.O., 2006. Artificial neural network model as a data analysis
tute of Technology, and Dr. Amor V.M. Ines, Department of
tool in precision farming. Transactions of the ASABE 49 (6),
Biological and Agricultural Engineering, Texas A&M Univer- 2027–2037.
sity. Ito, Y., Omatsu, S., 1997. Category classification method using a
self-organizing neural network. International Journal of
Remote Sensing 18 (4), 829–845.
references
Keiko, K., Pahari, K., Tamura, M., Yasuoka, Y., 1999. Feasibility
analysis for vegetation classification from time series NDVI
data with “Green Census” data. In: Proceedings of the Asian
Abenney-Mickson, S., Yomota, A., Miura, T., 1997. Water balance Conference on Remote Sensing, Hong Kong, November 22–25,
of field plots planted with soybean and pumpkin. 1999.
Transactions of the ASAE 40, 899–909. Lacoul, M., Samarakkon, L., Honda, K., 2000. Vegetation mapping
Barnes, R.J., Dhanoa, M.S., Lister, S.J., 1989. Standard normal in ganges river basin for global mapping project. In:
variate transformation and de-trending of near-infrared Proceedings of the Asian Conference in Remote Sensing,
diffuse reflectance spectra. Applied Spectroscopy 43 (5), Taipei, Taiwan, December 4–8, 2000.
772–777. Lejano, R.P., 2006. Optimizing the layout and design of branched
Bosdogianni, P., Petrou, M., Kittler, J., 1997. Mixed pixel pipeline water distribution systems. Irrigation and Drainage
classification with robust statistics. IEEE Transactions on Systems 20 (1), 125–137.
Geoscience and Remote Sensing 35, 551–559. Liu, W., Wu, E.Y., 2005. Comparison of non-linear mixture
Brivio, P.A., Maggi, M., Binaghi, E., 2003. Mapping burned surfaces models: sub-pixel classification. Remote Sensing of
in Sub-Saharan Africa based on multi-temporal neural Environment 94, 154–1154.
classification. International Journal of Remote Sensing 24 (20), Liu, W., Seto, K., Wu, E., Gopal, S., Woodcock, C., 2004. ART-MMAP:
4003–4018. a neural network approach to subpixel classification. IEEE
Chen, Z., Tsang, L., Kuga, Y., Chan, C., 1993. Active polarimetric Transactions on Geoscience and Remote Sensing 42 (9),
microwave remote sensing of corn and multiparametric 1976–1983.
retrieval of corn parameters with an artificial neural network. Lu, J., Kolaczyk, E.D., Gopal, S., 2003. Gaussian mixture
Microwave and Optical Technology Letters 6, 611–615. discriminant anlaysis and sub-pixel land cover
Chen, K.S., Yen, S.K., Tsay, D.W., 1997. Neural classification of characterization in remote sensing. Remote Sensing of
SPOT imagery through integration of intensity and fractal Environment 84, 550–560.
information. International Journal of Remote Sensing 18 (4), Malingreau, J.P., 1986a. Monitoring tropical wetland rice
763–783. production systems: a test for orbital remote sensing. In:
Defries, R.S., Townshend, J.R.G., 1994. NDVI derived land cover Remote Sensing and Tropical Land Management. John Wiley
classification at global scale. International Journal of Remote and Sons, New York, N.Y.
Sensing 15 (17), 3467–3486. Malingreau, J.P., 1986b. Global vegetation dynamics: satellite
Delfrate, F., Wang, L.F., 2001. Sun flower biomass estimation using observations over Asia. International Journal of Remote
a scattering model and a neural network algorithm. Sensing 7 (9), 1121–1146.
International Journal of Remote Sensing 22 (7), 1235–1244. McClendon, R.W., Hoogenboon, G., Seginer, I., 1996. Optimal
Diane, M.M., Kaminsky, E.J., Rana, S., 1995. Neural network control and neural networks applied to peanut irrigation
classification of remote-sensing data. Computers & management. Transactions of the ASAE 39 (1), 275–
Geosciences 21 (3), 377–386. 279.
Droogers, P., Bastiaanssen, W., 2002. Irrigation performance using Paudyal, G.N., Pandit, D.S., Goto, A., 1991. Optimization of design
hydrological and remote sensing modeling. Journal of of on-farm channel network in an irrigation area. Irrigation
Irrigation and Drain Engineering 128 (1), 11–18. and Drainage Systems 5 (4), 383–395.
Fisher, A., 1994. A model for the seasonal variations of vegetation Pons, L., 2008. Rice yield mystery solved? Key clues revealed about
indices in coarse resolution data and its inversion to extract declines in Asia. In: Agricultural Research, available at http://
crop parameters. Remote Sensing of Environment 48, 220–230. findarticles.com/p/articles/mi m3741/is 9 52/ai n6210362.
Foody, G.M., Cox, D.P., 1994. Sub-pixel land cover composition Accessed 30th January 2008.
estimation using a linear mixture model and fuzzy Roberts, D.A., Gardner, M., Church, R., Ustin, S., Scheer, G., Green,
membership functions. International Journal of Remote R.O., 1998. Mapping chaparral in the Santa Monica Mountains
Sensing 15, 619–631. using multiple endmember spectral mixture models. Remote
Foody, G.M., Lucas, R.M., Curran, P.J., Honzak, M., 1997. Non-linear Sensing of Environment 65, 267–279.
mixture modeling without end-members using an artificial Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973.
neural network. International Journal of Remote Sensing 18, Monitoring vegetation systems in the Great Plains with ERTS.
937–953. In: Proceedings of the 3rd ERTS Symposium, vol. 1, Greenbelt,
Haykin, S., 1994. Neural Networks: A Comprehensive Foundation. MD: NASA SP-351, pp. 48–62.
Macmillan College Publishing Company, New York, N.Y. Seckler, D., Baker, R., Amarasinghe, U., 1999. Water Security in the
Hunsaker, D.J., Barnes, E.M., Clarke, T.R., Fitzgerald, G.J., Pinter Jr., twenty first century. Water Resources Development 15, 29–42.
P.J., 2005. Cotton irrigation scheduling using remotely sensed Settle, J.J., Drake, N.A., 1993. Linear mixing and the estimation of
and FAO-56 basal crop coefficients. Transactions of the ASAE ground cover proportions. International Journal of Remote
48 (4), 1395–1407. Sensing 14, 1159–1177.
Ines, A.V.M., 2002. Improved crop production integrating GIS and Shahin, M.A., Tollner, E.W., Gitaitis, R.D., Sumner, D.R., Maw, B.W.,
Genetic Algorithm. PhD Dissertation. Asian Institute of 2002. Classification of sweet onions based on internal defects
Technology, Bangkok, Thailand. using image processing and neural network techniques.
Ines, A.V.M., Honda, K., 2005. On quantifying agricultural and Transactions of the ASAE 45 (5), 1613–1618.
water management practices from low spatial resolution RS Shimabukuro, Y.E., Smith, J.A., 1991. The least-squares mixing
data using genetic algorithms: a numerical study for models to generate fraction images derived from remote
76 c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 5 ( 2 0 0 9 ) 65–76

sensing multispectral data. IEEE Transactions on Geoscience Uchida, S., 2001. Sub-pixel classification of land use using
and Remote Sensing 29, 16–20. temporal profile of NDVI. Japan Society of Photogrammetry
Sugita, M., Yasuoka, Y., 1996. Land Cover Classification of East and Remote Sensing 40, 43–54.
Asia Using Fourier Spectra of Monthly NOAA AVHRR NDVI Van Dam, J.C., Huygen, J., Wesseling, J.G., Feddes, R.A., Kabat, P.,
Data. In: Proceedings of the IEEE International Geoscience and Van Waslum, P.E.V., Groenendjik, P., Van Diepen, C.A., 1997.
Remote Sensing Symposium, Lincoln, NE, USA, May 26–30, Theory of SWAP version 2.0: Simulation of water flow and
1996. plant growth in the soil–water–atmosphere–plant
SWAP, 2008. Soil Water Atmosphere Plant. Available at environment. Technical Document 45, Wageningen
http://www.swap.alterra.nl/. Accessed 30th January 2008. Agricultural University and DLO Winand Staring Centre, The
Takeuchi, W., Tamura, M., Yasuoka, Y., 2003. Estimation of Netherlands.
methane emission from West Siberian wetland by scaling Xiao, X., He, L., Salas, W., Li, C., Moore III, B., Zhao, R., Frolking, S.,
technique between NOAA AVHRR and SPOT HRV. Remote Boles, S., 2002. Quantitative relationships between
Sensing of Environment 85, 21–29. field-measured leaf area index and vegetation index derived
Tang, L., Tian, L., Steward, B.L., 2003. Classification of broadleaf from vegetation images for paddy rice fields. International
and grass weeds using gabor wavelets and an artificial neural Journal of Remote Sensing 23 (18), 3595–3604.
network. Transactions of the ASAE 46 (4), 1247–1254. Xiao, X., Boles, S., Frolking, S., Li, C., Babu, J.Y., Salas, W., Moore III,
Tateishi, R., Shimazaki, Y., Gunin, P.D., 2004. Spectral and B., 2006. Mapping paddy rice agriculture in South and
temporal linear mixing model for vegetation classification. Southeast Asia using multi-temporal MODIS images. Remote
International Journal of Remote Sensing 25 (20), 4203–4218. Sensing of Environment 100, 95–113.
Tedeschi, A., Menenti, M., 2002. Simulation studies of long-term Xu, M., Lu, J., Ruan, B., 2006. Crop Type change monitoring by
saline water use: model validation and evaluation of remote sensing in an irrigated district under water resources
schedules. Agricultural Water Management 54 (2), 123–157. restrictions. In: Proceedings of the World Environmental and
Tod, I.C., Wallender, W.W., Henderson, D.W., Devries, J.J., 1990. Water Resources Congress, Omaha, Nebraska, USA, May
Considerations for sizing water delivery systems. Irrigation 21–25, 2006.
and Drainage Systems 4 (2), 171–179.