Artif Life Robotics (2010) 15:221–224 DOI 10.1007/s10015-010-0797-4
© ISAROB 2010
ORIGINAL ARTICLE Takashi Yamaguchi · Kazuya Kishida · Eiji Nunohiro Jong Geol Park · Kenneth J. Mackin · Keitaro Hara Kotaro Matsushita · Ippei Harada
Artificial neural networks paddy-field classifier using spatiotemporal remote sensing data
Received and accepted: April 17, 2010
Abstract Monitoring changes in a paddy-field area is important since rice is a staple food and paddy agriculture is a major cropping system in Asia. For monitoring changes in land surface, various applications using different satellites have been researched in the field of remote sensing. However, monitoring a paddy-field area with remote sensing is difficult owing to the temporal changes in the land surface, and the differences in the spatiotemporal characteristics in countries and regions. In this article, we used an artificial neural network to classify paddy-field areas using moderate resolution sensor data that includes spatiotemporal information. Our aim is to automatically generate a paddy-field classifier in order to create localized classifiers for each country and region. Key words Artificial neural network · Classification · Remote sensing · MODIS
Monitoring a paddy-field area with remote sensing is difficult because a paddy has an annual cycle that can be classified into three main periods,1 (1) the flooding and rice transplanting period, when the land surface is observed as water, (2) the growing period, when an increasing vegetation index is observed, and (3) the fallow period, when the land surface is observed as soil. To monitor changes in land cover, moderate resolution remote sensing is effective because of the high frequency with which these satellites can scan the same area. In past research on paddy-field area estimation using remote sensing, decision trees or stochastic-analysis-based methods using spatiotemporal information were proposed.1,2 On the other hand, it is difficult to apply the same models to different countries and regions. Here, we applied an artificial neural network to classify a paddy-field area using moderate resolution remote sensing data in order to generate the classifier automatically.
2 Method 1 Introduction Monitoring changes in a paddy-field area is important since rice is a staple food and paddy agriculture is a major cropping system in Asia. For monitoring changes in land surface, many satellites have been launched and their applications were researched in the field of remote sensing.
2.1 Multilayered perceptron
J.G. Park · K. Hara · K. Matsushita · I. Harada Department of Environmental Information, Tokyo University of Information Science, Chiba, Japan
A multilayered perceptron (MLP) is a type of artificial neural network (ANN) that can approximate a complex function by machine learning. In this research, we used the MLP shown in Fig. 1 in order to develop a classification function for a paddy-field area from the MODIS data set. The MLP consists of three layers, the input layer with n neurons and a bias neuron, the hidden layer with m neurons and a bias neuron, and the output layer with K neurons. Each neuron is connected with every neuron in the next layer, and each connection has a weight value. When an input signal x = {x1,x2, . . . xn} is given, the j-th output signal zj of the hidden layer’s neuron and the k-th output signal yk of the output layer’s neuron are calculated by following expressions:
This work was presented in part at the 15th International Symposium on Artificial Life and Robotics, Oita, Japan, February 4–6, 2010
⎛ ⎞ zj = f ⎜ ∑ w ji xi ⎟ ⎝ i=0 ⎠
T. Yamaguchi (*) · K. Kishida · E. Nunohiro · K.J. Mackin Department of Information Systems, Tokyo University of Information Sciences, 4-1 Onaridai, Wakaba-ku, Chiba 265-8501, Japan e-mail:
[email protected]
n
(1)
222 Fig. 1. MLP network structure
⎛ m ⎞ yk = f ⎜ ∑ wkj zj ⎟ ⎝ j =0 ⎠
(2)
where i = 0, 1, 2 . . . , n, j = 0, 1, 2 . . . , m, k = 0, 1, 2 . . . , K, f is the activation function, and z0 and x0 are bias neurons. A bias neuron always outputs 1.0 to the next layer’s neurons. The sigmoid function was used for the activation function. MLP modifies each weight value using back-propagation (BP) training.3 Let xp = {x1p, x2p, . . . xnp}, where p = 1, 2, . . . , N is the p-th input signal, and tp = {t1p,t2p,…tKp } is the p-th teaching signal. The teaching signal is a true output signal that corresponds to the p-th input signal xp where the projection function can be defined as tp = g ( xp )
(3)
When the p-th training pattern {xp, tp} is given, the BP training modifies the weights for minimizing the mean square error E, defined as E=
1 N p ∑ t − yp N p=1
2
(4)
∂E + μ ⋅ Δw ji ( s − 1) ∂w ji ∂E Δwkj ( s ) = −λ + μ ⋅ Δwkj ( s − 1) ∂wkj
p ⎧ positive if y > θ Output p = ⎨ ⎩negative otherwize
(5)
(7)
where θ is a predefined threshold value. The MLP learning result is unstable from the initialization problem that MLP learning falls into different local minima according to the initial weight values. For resolving instability, a combination with ensemble learning and MLP is commonly used. Ensemble learning is a method for improving the stability of machine learning algorithms by using multiple learners. For the ensemble method, a bagging method was used.4 Bagging is a typical ensemble method that aggregates multiple training results. For aggregation, voting was used, as this is commonly used in bagging for classifiers. Let L(x) be an aggregated learner, Ls(x) be a multiple weak learner where s = 1, 2, . . . , r; and c = 1, 2, . . . , and C be a class label. A robust learner L(x) is defined by the expression L ( x ) = arg max {s; Ls ( x ) = c}
At the training step in BP training, the weight modifications Δwji(s) and Δwkj(s) are defined as Δw ji ( s ) = −λ ⋅
value, so that it is necessary to decide the positive or negative from the continuous output value. In this experiment, the p-th final output was defined by the function
c
(8)
Each learner is trained by using bootstrap samples.5 Let T be a training data set. Then the training data subset Ts ⊆ T for the s-th learner is constructed by using random samples.
(6)
where λ is a learning rate, and μ is a momentum rate. Each weight is commonly initialized by a random value. As a result of training, the MLP learns a function g(x) by modifying the weight values. In this research, the MLP was used as a two-class classifier which classifies positive or negative (1 or 0) for each paddy-field class. However, the MLP output is a continuous
3 Experiment 3.1 Paddy-field area estimation using moderate resolution remote sensing For this work, we used moderate resolution imaging spectroradiometer (MODIS) data collected at Tokyo University
223 Table 1. MLP parameters for the paddy classifier using MODIS sensor data Parametersa
3 3 2 2
A
C
B
D
bands bands bands bands
3 months 11 months 3 months 11 months
n
m
K
λ
μ
θ
9 33 6 22
9 33 6 22
1 1 1 1
0.25 0.25 0.25 0.25
0.05 0.05 0.05 0.05
0.3 0.3 0.3 0.3
n, input size; m, hidden size; K, output size; λ, learning rate; μ, momentum rate; θ, output threshold
a
red) data.6 From this result, we suggested that a MLP paddy-field classifier could not yield sufficient accuracy for practical use. To improve the classification accuracy, we investigate using three bands as input signals. In the paddy-field annual cycle, the features of paddy fields are the vegetation, soil, and water indices. For the vegetation index, NDVI was commonly used. NDVI is defined as NDVI = ( RED − NIR ) ( RED + NIR )
Fig. 2. NDVI maps and land-truth data for the north region of Kyushu, Japan
of Information Sciences, TUIS, Japan. TUIS receives satellite MODIS data from over eastern Asia, and provides these data for open research use. Figure 2 shows normalized difference vegetation index (NDVI) maps for the north region of Kyushu, Japan, in three different months (A, January; B, June; C, September) derived from 1-month composite MODIS sensor data. NDVI is a vegetation index defined by bands 1 and 2 (visible red and near infrared). Part D in Fig. 2 shows the land-truth data provided by the Japanese Ministry of Land, Infrastructure, Transport and Tourism (JMLIT). The light-colored pixels show that the paddy-area ratio is larger than other land-use types in the corresponding 500 m × 500 m area. The land-truth data are provided in vector data format, so the data were converted into raster format of 500-m scale pixel data (1 pixel = 500 m × 500 m resolution) in order to use the data as the solution set. From these maps, it can be seen that the vegetation in a paddy-field area changes over time. However, it is difficult to extract a generalized rule for paddy classification from this spatiotemporal information. Because the annual cycle of a paddy is different in each country and region, the changes in spatiotemporal information are also different in each region. Our aim is to automatically generate a paddy classifier using an artificial neural network and spatiotemporal MODIS sensor data shown in D of Fig. 2. 3.2 MLP paddy-field classifier In previous work, we evaluated a MLP paddy-field classifier using spatiotemporal MODIS bands 1 and 2 (red and infra-
(9)
where RED is visible red reflectance, and NIR is near infrared reflectance. Similarly, for indices of soil and water, a normalized difference soil index (NDSI) and a normalized difference water index (NDWI) were proposed by Takeuchi and Yasuoka.7 The NDSI and the NDWI are defined as follows: NDSI = ( SWIR − NIR ) ( SWIR + NIR )
(10)
NDWI = ( RED − SWIR ) ( RED + SWIR )
(11)
where SWIR is short wave infrared reflectance. Short wave infrared reflectance corresponds to band 6 data in the MODIS data set. Considering NDSI and NDWI, we also used band 6 data. In this work, we prepared four different MLP paddyfield classifier models, as shown in Table 1, for evaluating the improvement in accuracy using band 6 data. The differences for each model are input size n and hidden size m. In three-band models, 1 month of the input signals consist of band 1, band 2, and band 6. In two-band models, 1 month of the input signals consist of band 1 and band 2. In threemonth models, the input signals consist of band data from January, June, and September (2 or 3 × 3 inputs). These 3 months correspond to three periods of the annual paddy cycle. In 11-month models, the input signals consist of band data from January to November every separate month (2 or 3 × 11 inputs). The band data for each month were derived from 1-month composite MODIS sensor data of 500 m resolution. For the teaching signal, land-truth data were used. This value is either paddy or non-paddy (1 or 0), and was derived from digital national land information provided by the JMLIT. The parameters of MLP were defined by a previous experiment.
224 Table 2. Comparison of classification accuracy in the proposed paddy classifier Correct classification rate
3 3 2 2
bands bands bands bands
3 months 11 months 3 months 11 months
Total
Paddy
Non-paddy
0.879 0.908 0.873 0.876
0.746 0.714 0.719 0.707
0.908 0.953 0.908 0.915
3.3 Experimental results In this experiment, we evaluated the classification accuracy by using the proposed paddy classifier. To evaluate the classification accuracy, MODIS data were divided into two disjoint subsets, a training data set and a test data set, by using random sampling from the north region of Kyushu, Japan, as shown in Fig. 1. The number of test data sets was 10% of the number of training data sets. Table 2 shows the classification accuracy of the proposed paddy classifier. This table shows a comparison of the classification accuracy. It can be confirmed that the 3-band 11-month model yielded the best total classification rate. In addition, 3-band models yielded better results than 2-band models in total and paddy classification rates. This result shows the effectiveness of using band 6 data in the MLP paddy-field classifier. On the other hand, in the classification rate for paddy fields, 3-month models tended to yield a better result than 11-month models. This tendency was similar to that found in our previous research. It is expected that this reduction in accuracy was caused by the increasing input size. Considering the automatically generating paddy classifier, it is necessary to investigate feature selection methods because the annual cycle changes when the target region is changed.
4 Conclusions We have proposed a MLP paddy-field classifier using spatiotemporal MODIS band 1, band 2, and band 6 data in
order to generate a classifier automatically. From the computer simulation, we confirmed the improvement in classification accuracy by the additional use of band 6 data. This result shows the effectiveness of using band 6 data in the MLP paddy-field classifier. In addition, the proposed paddyfield classifier yielded a 0.908 classification rate. Considering that an accuracy of 0.95 or more is necessary for practical use, it can be concluded that the proposed MLP paddy-field classifier yields good result. In this experiment, we confirmed that an error occurred when converting from vector format into raster format for creating land-truth data. It is expected that the accuracy can be improved by reducing this error. In addition, we plan to compare this work with other paddy classifiers based on the decision-tree method, and other machine learning methods. Acknowledgment This research was supported by the research project of Tokyo University of Information Sciences for the sustainable development of economic and social structure dependent on the environment in eastern Asia.
References 1. Le Toan T, Ribbes F, Wang L, et al (1997) Rice crop mapping and monitoring using ERS-1 data based on experiment and modeling results. IEEE Trans Geosci Remote Sensing 1:41–56 2. Xiao X, Boles S, Frolking S, et al (2006) Mapping paddy rice agriculture in south and southeast Asia using multi-temporal MODIS images. Remote Sensing Environ 100:95–113 3. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning representations by back-propagating errors. Nature 323:533–536 4. Breiman L (1996) Bagging predictors. Mach Learn 24(2):123–140 5. Efron B, Tibshirani R (1993) An introduction to the bootstrap. Chapman and Hall, London 6. Yamaguchi T, Nunohiro E, Park JG (2009) Application of artificial neural network for land cover classification using spatiotemporal information. 5th International Conference on Information 7. Takeuchi W, Yasuoka Y (2004) Development of normalized vegetation, soil and water indices derived from satellite remote sensing data (in Japanese). J Jpn Soc Photogrammetry Remote Sensing 43(6):7–19
