Intelligent Systems for Agriculture in Japan - IEEE Control Systems ...

SPECIAL SECTION

By Yasushi Hashimoto, Haruhiko Murase, Tetsuo Morimoto, and Toru Torii

T

©DIGITAL STOCK 1996 and ©DIGITAL STOCK 1997

his article describes some findings of agricultural research studies being conducted in Japan in three areas: 1) artificial intelligence (AI) or computational intelligence applications in agriculture and the environment, 2) intelligent environment control for plant production systems, and 3) intelligent robots in agriculture. First, the latest biosystem-derived algorithms are discussed. A finite element inverse technique using a photosynthetic algorithm (PA) is described, followed by a comparison of neural network training by a photosynthetic algorithm versus a genetic algorithm (GA). Leaf cellular automata (LCA) are introduced, and their application to optimization problems is discussed. Second, a decision system consisting of neural networks (NNs) and GAs is applied to the optimization of plant growth under hydroponics in Japanese plant factories. In this system, plant growth as affected by the nutrient concentration is first identified using NNs, and then the optimal l-step set points of the nutrient concentration that maximize the plant growth are determined through simulation of the identified NN model using GAs. Finally, recent developments of intelligent agricultural robots in Japan are described. Much research on Hashimoto ([email protected]) and Morimoto are with the Department of Bio-mechanical Systems, Ehime University, Tarumi, Matsuyama 790-8566, Japan. Murase is with Osaka Prefecture University, Sakai, Japan. Torii is with the University of Tokyo, Tokyo 113-8656, Japan.

October 2001

0272-1708/01/$10.00©2001IEEE IEEE Control Systems Magazine

Authorized licensed use limited to: University of Michigan Library. Downloaded on May 19,2010 at 20:48:56 UTC from IEEE Xplore. Restrictions apply.

71

automation in agriculture has also been conducted in universities. Due to limited funding, most of this research has covered methodologies such as navigation, sensing, and the application of control theory. At research institutes and manufacturers, which have greater financial resources, more practical systems have been tested. Researchers are now integrating the new technologies for autonomous navigation in the field. It can be concluded that intelligent approaches are useful tools for mechanizing complex agricultural systems.

rived algorithms (BDAs) are suggested. Photosynthesis is one of the most important biochemical phenomena and can be viewed as a natural implementation of an optimization. It is sometimes mentioned that a plant is not optimized by nature to function as an energy conversion device due to its low energy conversion efficiency of about 3%. Comparisons are made to man-made devices such as photovoltaic cells and photoelectrochemical cells, which transform the sun’s energy into an electrical current or chemical fuels with efficiencies as high as 25%. This is an unfair comparison, however, because plants are under heavy functional constraints to maintain the diverse set of biological activities necessary for their survival and for the preservation of their species. A better comparison would require that only those biochemical pathways in the plant directly related to energy conversion be considered when calculating energy conversion efficiency. GAs have been extensively used in the controls field. However, the GA does not have to be the ultimate optimization technique. Two different biosystem-derived optimization algorithms using the mechanism of the photosynthetic pathway were developed by one of the authors. The BDAs referred to are the PA and the LCA. In the following sections, the principles of the PA and LCA are introduced briefly. The developed biosystem-derived optimization algorithms contribute to control applications by providing additional search technique options.

Photosynthesis is one of the most important biochemical phenomena and can be viewed as a natural implementation of optimization. Computational Intelligence in Agriculture and the Environment Many problems in agricultural engineering involve optimizing different types of biosystems, such as drainage and irrigation systems, crop scheduling, and the handling and blending of materials. Such biosystems typically depend on decision parameters that can be chosen by the system designer or operator. An inappropriate choice of decision parameters causes serious flaws in performance, as measured by some objective or fitness function. Another problem often encountered in agricultural engineering involves testing and fitting of quantitative models. Engineering or scientific research in any problem area classically consists of an iterative process of building explanatory or descriptive models, collecting data, testing the models, modifying the models when discrepancies are found, and then repeating the process until the problem is solved satisfactorily. The problems that deal with optimizing biosystems and fitting quantitative models eventually require refinement or processing using adaptive search procedures or optimization techniques. There are many search techniques, including exhaustive techniques (random walk), calculus-based techniques (gradient methods), partial knowledge techniques (hill climbing), knowledge-based techniques (production rule systems, heuristic methods), stochastic techniques (simulated annealing), and biologically inspired algorithms (genetic and immune system algorithms). In realistic systems, the interactions between the parameters are not generally amenable to analytical treatment, and researchers must resort to appropriate search techniques. Recently, genetic and immune system algorithms have received considerable attention due to their ability to locate very good solutions in extremely large search spaces with reasonable computational effort [1], [2]. It is interesting to note that if one looks carefully at plant systems or phytosystems, many different biosystem-de-

72

Principle of the Photosynthetic Algorithm In the diagram of the Benson-Calvin cycle [3] in Fig. 1(a), each line represents the conversion of one molecule of each metabolite. Fig. 1(a) also indicates that the product of the Benson-Calvin cycle is DHAP (dihydroxyacetone-P). Some of the DHAP, which may be unstable and/or low quality, is reused or reprocessed in the cycle. The remaining portion of the product DHAP, which may be stable and/or high quality, stays as starch. The refined high-quality DHAP can be considered the knowledge string that is conceptually equivalent to the final form of the chromosome (solution or estimate) in a GA. The Benson-Calvin cycle includes many different recombinations of molecules that are again conceptually equivalent to the crossover of chromosomes in the GA. In the PA, the crossover is regulated more strictly by the photosynthetic rules than the crossover operator of the GA. Fig. 1(b) shows the part of the photorespiratory system that contains the Benson-Calvin cycle. The biochemical balance between the Benson-Calvin cycle and photorespiration can be viewed as a natural implementation of an optimization procedure that maximizes the efficiency of sugar production under the continuously variable energy of the sun. The PA utilizes this unique natural optimization process, which is analogous to the mutation operator in the GA.

IEEE Control Systems Magazine

October 2001


3 ATP

Ribulose-5-P

3 ADP Ribose-5-P

Ribulose-1, 5-P2 (RuBP)

Xylulose-5-P

3 CO2 GAP

Sedoheptulose-7-P Pi H2O Sedoheptulose-1, 7-P2

Glycerate-3-P 6 ATP 6 ATP

DHAP

Erythrose-4-P

Xylulose-5-P Glycerate-1, 3-P2 Fructose-6-P Pi

GAP

6 NADPH 6 NADP

H2O Fructose-1, 6-P2

6Pi GAP

DHAP

GAP

DHAP (Product or Feedback) (a) Chloroplast

Glycerate

BensonCalvin Cycle

ADP

AT

Peroxisom Glycolate Pi

DHAP

Glycolate-2-P

O2

product is evaluated based on a fitness value, as obtained by calculating the difference between the output value of the system using parameters currently given by the PA and the training output data. Fig. 3, which presents the whole process given in Fig. 1(a), shows a flow diagram indicating the calculation process of the PA. The process begins with the random generation of light intensity. CO2 fixation rate is then evaluated by (1) based on light intensity. Depending on the fixation rate, either the Benson-Calvin cycle or the photorespiration cycle is chosen for the next process. In both cycles, 16-bit strings are shuffled according to the carbon molecule’s recombination rule in photosynthetic pathways. After some iteration, GAPs, which are intermediate knowledge strings, are produced. Each GAP consists of 16 bits. The fitness of these GAPs is then evaluated, and the best fit GAP remains as a DHAP (current estimated value). One of the unique features of this algorithm is the inherent stimulation function. The stimulation occurs due to randomly changing light intensity, which alters the degree of influence on renewing the elem e n t s o f R u B P b y p h o t orespiration. The frequency of the stimulation cycle by photorespiration can be calculated by the CO2 fixation rate given by

(b)

Figure 1. (a) Photosynthetic pathways of Benson-Calvin cycle and (b) photorespiration. The PA uses rules governing the conversion of carbon molecules from one substance to another in the Benson-Calvin cycle and photorespiration reactions. Fig. 2 illustrates the variation of recombination of carbon molecules appearing in the PA. For example, RuBP (rubilose-biphosphate) consists of three sets of a 5-carbon-molecule substance that react with a 3-carbon-dioxide molecule to produce five sets of a 3-carbon-molecule substance (GAP). The product of photosynthesis, DHAP, as shown in Fig. 1(b), provides the knowledge strings of the algorithm. Optimization is attained when the quality of a product no longer improves. The quality of a

October 2001

C=

Vmax 1+ A/ L

(1)

where C = CO2 fixation rate, Vmax = maximum CO2 fixation rate, A = affinity of CO2, and L = light intensity. The parameters involved in (1) are all determinable, but their values can be assigned within a realistic range and need not be empirical. When executing the PA, the light intensity should be generated randomly. Alternatively, the actual light intensity varying with time through a measuring system may be used. Variation of the light intensity as a stimulant is effective in reducing the occurrence of local minima traps in the search procedure. The CO2 concentration in the leaf varies depending on the CO2 fixation rate. The ratio of O2 concentra-



73

RuBP

RuBP

+ Oxygen

3CO2 Ribose-5-P

+

Sedoheptulose-7-P Glycerate

GAP

Glycolate

Glycolate

Glycerate

DHAP Xylulose-5-P

(Benson-Calvin Cycle)

Fructose1, 6-P

Erythrose-4-P

GAP CO2

(a)

(b)

Figure 2. (a) Recombination of carbon molecules in the B-C cycle and (b) photorespiration.

Atmosphere Light (Stimulation)

Oxygen/CO2 Concentration

CO2 Reservior

RuBP BensonCalvin Cycle

PhotoRespiration

GAP

Fitness Copy Next Iteration

GAP

GAP

Poor

Discard

Good DHAP (Knowledge String)

Figure 3. The photosynthetic algorithm (PA). tion to CO2 concentration is evaluated to determine the ratio of the calculation frequency of the Benson-Calvin cycle to that of the photorespiration cycle.

Application of PA to Finite Element Inverse Analysis The finite element method is a powerful numerical procedure for solving mathematical problems in engineering and physics. The finite element method was also employed to devise a spatial NN in a control application [4]. Fig. 4 illustrates a possible application of the finite element NN in a plant growth control system. The learning algorithm of the finite element NN is an inverse solution of a problem described by Poisson’s equation. This finite element inverse problem is simply an optimization problem or parameter es-

74

timation problem. Calculus-based estimation techniques such as the least-squares and conjugate gradient are available for solving finite element inverse problems, and GAs or other BDAs may also be used. In this section, the performance of the PA in solving a finite element inverse problem is discussed. A cantilever beam system (5 unit length × 10 unit length) is provided as a numerical example (see Fig. 5). This finite element model consists of four linear triangular elements. The elastic properties (Young’s modulus and Poisson’s ratio) of each element are assumed to differ. Nodes 1 and 2 are fixed; the remaining nodes can be displaced freely. A unit vertical load is applied at node 4, and the displacements of nodes 3, 4, and 5 in the horizontal and vertical directions due to the unit load are observed. The PA is expected to search for the optimum values of the eight unknown elastic moduli, which are the Young’s modulus and Poisson’s ratio of each finite element. The parameters used for this test were as follows: the affinity of CO2 was set to 10,000; the maximum light intensity varied from 10,000 to 50,000 lux (an increase in light intensity implies an increasing likelihood for activating photorespiration); the maximum CO2 fixation rate was 30 -2 -1 mg⋅m ⋅s ; and the maximum number of cycles for the Benson-Calvin cycle and photorespiration were 30 and 45, respectively, per search iteration. In the PA procedure, the finite element evaluation appears in the fitness check process of the knowledge strings (DHAP). Each of the eight elastic moduli is coded in a 16-bit DHAP molecule. After converting them to decimal numbers, the nodal displacements at nodes 3, 4, and 5 are calculated using the estimated elastic moduli and the given boundary conditions. The observed displacement data are compared with the calculated output data to determine the fitness of the estimates. When a set of estimates with better fitness than the previous data is obtained, the best data set is stored in the DHAP reservoir for the next comparison. After


October 2001


completing one cycle of the search process with a predetermined freZ –1 u ( t − 1) quency for the Benson-Calvin cycle and photorespiration, the process Temperature Current Growth then randomly generates the light inHumidity y ( t ) y (t ) D tensity for the next iteration with a reVegetables Carbon Dioxide newed photosynthetic frequency Light u(t) condition. Root Water Potential The observed displacements at nodes 3, 4, and 5 are indicated in Table 1. Finite Element Structured Neural Network Growth Model The negative values of Y imply that nodes y(t−1) 3, 4, and 5 are displaced downward by Z –1 the force applied at node 4. In this simulaGrowth Indices Growth tion, the observed displacements were Status calculated from the predetermined elastic moduli. These predetermined elastic values are intended to be the target values in the estimation test. The estimation of the elastic moduli of the finite element model using the PA Figure 4. Conceptual representation of a finite element NN application to a plant growth was very satisfactory. Fig. 6 shows the control system. y is the desired growth rate. D convergence property of the fitness. A dramatic decline in the error level down to 10-4 was observed in the initial ten iterations. After Load 1,000 iterations, the PA converged to a total absolute error -4 level of about 2.8 × 10 . No significant improvement was observed after 200 iterations. Tables 2 and 3 summarize the 4 1 comparison of the estimated values of the elastic moduli (Young’s modulus and Poisson’s ratio) and corresponding 4′ 4 target values. The tables show that most of the estimated val5 ues are very close to their target values except the Poisson’s Fixed 3 1 ratio of element number 4. 5′

2

2

3

Principle of Leaf Cellular Automata

October 2001

Deformed

3′

Figure 5. Finite element model (cantilever beam).

8 7 6

Error ×10−4

5 4 3 2

920

990

780

850

710

570

640

430

500

290

360

150

220

0

10

1

80

LCA are one form of cellular automata and are an estimation engine that utilizes the rules of the photosynthetic pathways. LCA are derived from the interaction of substances on a leaf. They consist of two layers: the “surface layer” and the “inside layer,” as shown in Fig. 7. The surface layer has four elements: light, stoma, CO2, and O2. The inside layer has one element, starch, and represents a solution. LCA have two procedures: photosynthesis and photorespiration. Photosynthesis occurs when the conditions of the elements making up the surface layer are satisfied. Light, CO2, and stoma are needed and consumed. O2 is produced. Photorespiration occurs when the O2 concentration exceeds a certain value. Photorespiration consumes starch in the inside layer (strictly speaking, photorespiration consumes glycolate) and O2 on the surface layer. CO2 is emitted from the surface layer. In a minimum searching problem, each row represents a bit string of a solution. Therefore, if a cell space is 8 × 8, the inside layer has eight solutions that are expressed in 8-bit binary code. If the best solution is generated, the row of the surface layer that generated this solution is copied to the next surface layer.

Iteration Number

Figure 6. Convergence of estimates using the PA.



75

Table 1. Observed displacements at nodes 3, 4, and 5. Node

Table 2. Estimated values of Young’s modulus.

Displacement X

Element Y

Estimated

Target

1

116

120

3

−0.0332

−0.123

2

80

80

4

0.4140

−1.303

3

95

90

5

0.0011

−0.055

4

68

70

Application of LCA to a Search Problem The objective function indicated in Fig. 8 has many local minima. A minimum search test was conducted using this multipeak function. The convergence properties of a standard benchmark GA and LCA were compared. Fig. 9 shows that LCA performed better in terms of success rate. The success rate was calculated by dividing the number of times the minima were successfully found by the number of search trials. When the GA was tested, the solutions sometimes fell into local minima, so the percentage of success at 200 iterations is only about 50%. LCA found solutions more than 90% of the time in 200 iterations.

Intelligent Environment Control for Plant Production Systems Hydroponic culture techniques have several potential advantages over soil culture techniques for cultivation (e.g., flexible control of the root-zone environment) and for the mechanization of cultivation processes [5]. In hydroponic cultivation, therefore, the development of a more effective control technique will provide remarkable progress in plant production [6]. For the effective control of plant production, it is efficient to monitor the current physiological status of the plant and then use this information for control. Such an approach is known as the “speaking plant approach (SPA),” where the environmental factors are considered to be the input and the plant responses the output [7]. Generally, however, it is very difficult to control the plant responses because the physiological processes are quite complex and uncertain. Intelligent control approaches are more suitable than traditional mathematical methods for dealing with complex systems such as cultivation systems [8], [9]. NNs have the

Table 3. Estimated values of Poisson’s ratio. Element

Estimated

Target

1

0.24

0.25

2

0.30

0.35

3

0.29

0.30

4

0.23

0.32

capability to identify unknown complex systems with their own learning ability [10]. GAs are one combinatorial optimization technique. Using a multipoint search procedure, they search for an optimal value of a complex objective function by simulating the biological evolutionary process based on crossover and mutation in genetics [2], [11]. This section describes the application of a new intelligent control system consisting of a decision system and a feedback control system to optimize plant growth in hydroponic tomato cultivation. The decision system, which consists of NNs and GAs, provides the optimal set points of the nutrient concentration to maintain a balance between vegetative and reproductive growth. The control input is the nutrient concentration of the solution, and the controlled output is the growth of a tomato plant.

Plant Growth Optimization Problem The plant material used here is tomato plant (Lycopersicon esculentum Mill. cv. Momotaro) in hydroponics. In tomato cultivation, good fruit yield requires an optimal balance between 4

3

2

1

0.5

(a)

(b)

1.5

2

2.5

3

Objective Function

Figure 7. Two-layered cellular automata. (a) Surface layer; (b) inside layer.

76

1

Figure 8. A multipeak function used for the minimum search test.


October 2001


vegetative growth (e.g., root, stem, and leaf growth) and reproductive growth (e.g., flower and fruit growth). In hydroponic cultivation, however, vegetative growth becomes more active than reproductive growth because the roots of the plants are always in a suitable environment for the uptake of nutrient ions. Active vegetative growth induces poor reproductive growth. The nutrient concentration of the solution in hydroponics is one of the most important manipulated factors for adjusting the balance between the two types of growth [12]. It is usually increased with plant growth. The balance between the two types of growth is determined and fixed at the seedling stage. This means that optimal control during this stage is important. During the seedling stage, however, since only such vegetative growth as stem, leaf, and root growth can be observed, the future balance between the two types of growth has to be predicted from this growth data. It has been reported that the ratio of stem dry weight to root dry weight (S/R) is a good indicator for predicting the future balance and that a smaller value results in better yields [13]. It is well known that active stem growth is detrimental to reproductive growth in the cultivation of fruit and vegetables. In this study, leaf growth was adopted as one of the predictors instead of root growth because it can be measured using an image-processing system, and a larger leaf growth is linked to the promotion of photosynthate production in plants. From these findings, the ratio of total leaf length (TLL) to stem diameter (SD) was defined as a predictor for future balance in growth. Actually, higher values of TLL/SD resulted in better reproductive growth. Therefore, controls for maximizing TLL/SD may be valuable only during the seedling stage. Let TLL( k )/SD( k ) be a time series of TLL/SD as affected by nutrient concentration NC( k ) ( k = 1,..., N : sampling day, N: final day). For implementation, the seedling stage(1 ≤ k ≤ N ) was divided into four steps: 1) transplanting, 2) vegetative growth after transplanting, 3) flowering of the first truss, and 4) fruit setting for the first truss and flowering for the second truss, and the values of TLL( k )/SD( k ) at the last step (step 4) were evaluated. The value of the nutrient concentration in each step, NC1 , NC2 , NC3 , or NC4 , was kept constant [1 ≤ k ≤ N 1L (step 1), N 1L + 1 ≤ k ≤ N 2L (step 2), N 2L + 1 ≤ k ≤ N 3L (step 3), N 3L + 1 ≤ k ≤ N (step 4), where N 1L , N 2L , N 3L , and N represent the last days of the first, second, third, and fourth steps]. The objective function was given by the average value of TLL/SD at the last step (step 4, N 3L + 1 ≤ k ≤ N ) in its dynamic response as follows (N 3L + 1: first day of step 4): F ( NC) =

1 N − N 3L + 1

N

TLL( k ) . k = N 3 L + 1 SD( k )

∑

(2)

Thus, the optimization problem is to determine the optimal four-step set points of nutrient concentration, NC1 , NC2 , NC3 , and NC4 , which maximize F ( NC). The nutrient concentration here was constrained to 0.2 ≤ NC( k ) ≤ 20 . (mS/cm). maximize F ( NC) subject to 0.2 ≤ NC( k ) ≤ 20 . (mS/cm).

October 2001

100 80 60

LCA

%

GA

40 20 0 0

50

100

150

200

Iteration Number

Figure 9. The convergence properties of GA and LCA.

Design of a Control System for Optimization Fig. 10(a) shows the schematic diagram of a control system for a deep hydroponic system. The plant roots in the deep hydroponic system always dip into the nutrient solution. The nutrient concentration of the solution is automatically adjusted to the appropriate set point by mixing the highly concentrated nutrient solution and water. For growth optimization of plants in plant factories, optimal control of the environment is essential and requires taking the physiological status of the plant into consideration. As mentioned earlier, this has been explored as an SPA [7]. Fig. 10(b) shows the block diagram of a control system based on the SPA. It consists of a decision system for determining the optimal set points of the environment and a feedback control system for maintaining the environment at the optimal set points [14]-[16]. A decision system, which consists of NNs and GAs, determines the optimal set points of the environment on the basis of plant growth data. In this method, plant responses affected by environmental factors are first identified using NNs, and then the optimal environmental set points are searched for through simulation of the identified NN model using GAs. The study found that the time variation in the physiological dynamics of the plant, along with the growth, could be captured by using a recurrent identification and search technique to determine the optimal values. That is, the identification of plant responses and the search for optimal values are periodically repeated to follow changes in the physiological dynamics of the plants. The optimal value can be changed according to the change in the physiological status of the plant.

Neural Networks In the study, NNs were used for creating black-box models for simulation, which predict the TLL/SD ratio of the nutrient concentration of the solution. For dynamic identification, arbitrary feedback loops that produce time histories of



77

the data are necessary elements of the network [17], [18]. Fig. 10(c) shows a time-delay NN used for identifying the response of the TLL( k )/SD( k ) ratio to two inputs: nutrient concentration, NC( k ), and light intensity, L( k ) [19]. The current output TLL( k )/SD( k ) is estimated from both the historical input data {NC( k ),..., NC( k − n), L( k ),..., L( k − n)} and the historical output data {y( k − 1),..., y( k − n)} (n: system order). The learning method was error back-propagation [20]. The data samples are divided into two data sets: a training data set and a testing data set. The former is used for training the NN, and the latter for evaluating the accuracy of the identified model. This type of model validation is

called “cross validation.” The system order and number of the hidden neurons in the NN were determined based on cross validation.

Genetic Algorithms To employ GAs, an “individual” for genetic evolution must first be defined. Fig. 10(d) illustrates the definition of individuals and population P( t ) used in the GA application. Since the purpose is to determine the four-step set points of the nutrient concentrations that maximize F ( NC), the set points NC1 , NC2 , NC3 , and NC4 represent an individual and each nutrient concentration is coded as a 6-bit binary string (e.g.,

Control System

Decision System New Set Point

Search for Optimal Set Points

Plants

Identification Neural Networks

Genetic Algorithms

Control Device Highly Concentrated Water Nutrient Solution

Sensor Set Point

Mixing Tank

e

Feedback Controller

+ −

Environment

Plant

Hydroponic System (a)

(b)

Population P(t) at Generation t Individual 1

NC(k) Time Series of Nutrient Concentration

Time Series of Light Intensity

Past Time Series of TLL/SD

NC1 NC2 100110 001100

NC3 NC4 000111 101010

. . . . . . . . . . . . . . . .

NC(k−n)

y(k) =

L(k)

Individual N

000111 111100

001111 000010

TLL/SD Genetic Operators

Output

Crossover Mutation Selection

L(k−n) y(k−1)

New Population P(t + 1) at Next Generation (t + 1) Individual 1

y(k−n)

010111 100001 001111 110010

. . . . . . . . . . . . . . . .

Time Delay

Individual N

(c)

100110

111001 101100 100010 (d)

Figure 10. Schematic diagrams of (a) a deep hydroponic system; (b) a control system consisting of a feedback control system and a decision system; (c) an NN used in the decision system; (d) a GA used in the decision system.

78


October 2001


350

250

Treat. 1

350

Treat. 2 Treat. 3

300

200

TLL/SD

TLL/SD

300

150 100

Estimated 0

5

10 (a)

15

Observed

20

100 50

20

Light Intensity × 102(µE m−2)

200 150

50

0

15

5

10

15

20

Time [Days]

10

Figure 12. A comparison of the estimated response and the observed response of the TLL/SD ratio.

5 0 0

Nutrient Concentration (mS/cm)

250

5

10

15

20

An elitist strategy was used for selection (i.e., the best individual in a generation was carried through to the next generation).

1.5

Measurement and Identification of the TLL/SD Ratio

1 0.5 0

0

5

10

15

20

Time [Days] (b)

Figure 11. The observed daily changes in (a) the TLL/SD ratio of tomato plants and (b) the light intensity and nutrient concentration of the solution during the seedling stage. individual i = NCi 1 , NCi 2 , NCi 3 , NCi 4 = 100100, 001001, 001100, 101010). A new population P( t + 1) is generated through crossover, mutation, and selection (t: generation index). Fitness, which is given by (2), is an indicator for measuring an individual’s survival quality. All individuals are evaluated based on their fitness values. During the evolution process, individuals having higher fitness reproduce and individuals with lower fitness die in each generation. An individual having the maximum fitness is regarded as an optimal solution. The procedure of the GA we have employed is as follows. • Step 1: An initial population consisting of several individuals is generated at random. • Step 2: Several individuals in another population are added to the original population to maintain diversity. • Step 3: Crossover and mutation operations are applied to the individuals selected at random. • Step 4: The fitness values of all individuals are calculated using the NN model, and their performances are evaluated. • Step 5: Superior individuals are selected and retained for the next generation (Selection). • Step 6: Steps 2 through 5 are repeated until an arbitrary condition is satisfied.

October 2001

First, the data for identification were obtained. Fig. 11 shows the daily changes in the TLL/SD ratio observed for tomato plants grown in hydroponics, as well as the light intensity and nutrient concentration of the solution during the seedling stage. The control period is restricted to the seedling stage. Three patterns of the TLL/SD ratio under three different nutrient concentration treatments are shown. Morimoto et al. found that three or more data sets are necessary for identification [21]. These data were measured every day using an image-processing unit and a ruler. The light condition was arbitrary. The value of the TLL/SD ratio was found to be markedly affected by nutrient concentration. For identification, the data for N = 22 were obtained in each pattern. The response of the TLL/SD ratio to both nutrient concentration and light intensity was then identified by an NN, and a black-box model was created for predicting the TLL/SD ratio. Fig. 12 shows the identification result in the response of the TLL/SD ratio to both light intensity and nutrient concentration. The data used here were independent of the data in Fig. 11. To save computing time, n = 1 was selected as the system order. It was also found that the number of hidden neurons N h = 5 was best for cross validation. The estimated responses were closely related to the observed responses, which means that a reliable computational model could be obtained for predicting the behavior of the TLL/SD ratio under any combination of the four-step set points of nutrient concentration.

Search for the Optimal Set Points of Nutrient Concentration Fig. 13 shows an evolution curve during the search for an optimal value under different crossover and mutation rates. The fitness in all cases dramatically increased and then reached a maximum value. However, the degree of increase



79

Optimized Control Performance of the TLL/SD Ratio Fig. 14 shows the actual control performance of the TLL/SD ratio. The solid line represents the optimized control performance, and the dotted line represents the conventional control performance. The conventional strategy is simply to increase the nutrient concentration in a stepwise fashion with the growth of the plants. To clarify the difference between the two control performances, standard deviations were calculated and a t-test was then carried out. Comparing both control performances, it is apparent that the values of 200

TLL/SD

150

Pc = 0.2, Pm = 0.02

300

Pc = 0.2, Pm = 0.2 Pc = 0.8, Pm = 0.8

0

10

20

30

100 Conventional Control

Stem Diameter (cm)

0

1 0.8 0.6 0.4 0.2 0

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

0

5

10

15

20

25

140 120 100 80 60 40 20 0

2.5 2 1.5 1 0.5 0

Time [Days]

Generation Number

Figure 13. Evolution curves of the search for an optimal value under different crossover and mutation rates.

80

Optimized Control

50

Nutrient Concentration (mS/cm)

Fitness

310

290

ing of the first and second trusses and the fruit setting of the first truss during the seedling stage).

Total Leaf Length (cm)

is larger for higher crossover and mutation rates than for lower crossover and mutation rates. For example, the fitness reached a maximum value at the ninth generation when the crossover and mutation rates were high (Pc = 0.8 and Pm = 0.8). When the crossover and mutation rates were low (Pc = 0.2 and Pm = 0.02), however, the fitness could not reach the maximum value and fell into a local optimum. This is probably due to the loss of diversity in the population caused by low crossover and mutation rates. Note, however, that there is no guarantee that GAs yield a global optimal solution. In this study, an optimal value obtained from a GA was confirmed by using a round-robin algorithm that systematically searches for all values (possible solutions) around the optimal solution at the proper step. An optimal solution was also confirmed with a different initial population and different methods of crossover and mutation. The optimal four-step set points of the nutrient concentration obtained from the decision system were a slightly higher level (1.4 mS/cm) in the first step, a markedly lower level (0.3) in the second step, a slightly higher level (1.6) in the third step, and the maximum level (2.0) in the fourth step. In hydroponics, as mentioned above, since the roots of plants are always in a suitable environment for the uptake of nutrient ions, vegetative growth during the seedling stage is easy to promote. Active vegetative growth during the seedling stage will result in poor reproductive growth in the future [12], [22], [23]. Therefore, vegetative growth must be suppressed at the early seedling stage, before the flowering of the first truss. The low nutrient concentration in the second step seems to be effective in suppressing excessive vegetative growth during the seedling stage. The high nutrient concentrations in the third and fourth steps appear to be useful in accelerating reproductive growth (i.e., the flower-

Figure 14. Performance with optimized and conventional control.


October 2001


Table 4. The number of presentations on agricultural robots at the annual meeting of JSAM. Year

1991

1992

1993

1994

1995

1996

1997

1998

1999

2000

Number of presentations

25

20

31

34

15

37

33

40

34

32

the TLL/SD ratio are 10-15% higher with the optimal control than with the conventional control. This result was confirmed using a t-test at the 5% level of significance. With the optimal control, the reason for this result is that stem growth was significantly suppressed by the low nutrient concentration at the second step, whereas the leaf growth did not vary significantly in either case. Thus, the effectiveness of this control system was also confirmed experimentally.

Intelligent Robots in Agriculture

the surrounding soil area, detection of boundary lines between crop and soil areas, and position identification using a three-dimensional perspective view transformation are required. Discrimination of crop area was performed using color transformation of an HSI (hue, saturation, and intensity) transform [24]. Fig. 15 shows the result of the HSI transfer of cloudy and sunny day images taken at 12:00 p.m. Discrimination between the crop canopy and soil area was successful using the HSI transfer without the influence of climate and shooting time. A least-squares method was used for boundary detection between crop row and soil area, and a three-dimensional perspective view transformation was used for position identification. The results showed that the offset error was within 0.02 m and the attitude angle error was within 0.5°, which were sufficient for guidance in the field. This algorithm was applied to a vision-guided tractor [25]. Fig. 16 shows the resulting path trajectory and that the offset error was within 0.02 m at a speed of 0.25 m/s. Work is continuing on this project to increase the speed, and work on vision guidance in the paddy field is in progress. At Hokkaido University, an NN vehicle controller was designed in which the motion of a mobile agricultural robot was specified as a nonlinear system with high learning ability [26]. At Kyoto University, an automatic “follow-up vehicle,” using two small head-feeding combines, is under development [27]. A human operator in the front vehicle controls it, and the follow-up vehicle is automatically controlled by computer. At

In Japan, agricultural robotic research is widely performed in the areas of autonomous navigation, harvesting, and nursery production. Table 4 shows the number of presentations on robotic research at the annual meetings of the Japanese Society of Agricultural Machinery (JSAM) over the past ten years. Research in autonomous navigation is being conducted in universities, in government institutes, and by agricultural machinery manufacturers. In universities, due to financial limitations, most of the research has focused on methodologies such as navigation, sensing, and the application of control theory. At research institutes and manufacturers, which have more financial resources, more practical systems were developed. Research in harvesting robots is performed mainly in universities, though the technical levels are still beneath that of the MAGARI robot, which was developed at CEMAGREF (France) in the late 1980s. Nursery robots are developed mainly by government research institutes and manufacturers, and some of them are reaching the market. In particular, grafted nurseries, such as cucumber, watermelon, tomato, and eggplant, are widely used in greenhouses, and various types (a) (b) of grafting robots are being developed by agricultural manufacturers Figure 15. Results of HIS transformation. (a) Cloudy day (12:00); (b) sunny day (12:00). and other types of industries. 0.6

At the University of Tokyo, a machine vision algorithm for crops was developed and applied to vision-guided navigation of a tractor, which would be used for row crop husbandry such as mechanical weeding and precise chemical applications. For vision guidance, image analysis of the crop field is essential. Thus, highly accurate discrimination of crop area from

October 2001

0.4

0.8 Target Line

0.4 0

Offset (m)

Autonomous Navigation

Offset (m)

1.2

0

5

10

15 Trace

−0.4

Target Line

0.2 0 −0.2

Trace

−0.4 −0.6

0

5

10

Distance [m]

Distance [m]

(a)

(b)

15

20

Figure 16. Results of vision-guided navigation. (a) Trace on an artificial lawn; (b) trace on a crop row. IEEE Control Systems Magazine


81

Reflectance [%]

work was completed within 2 hours and 15 minutes at a speed of 0.45 m/s. The work area was 50 × 100 m2. A self-diagnosis function and an alarm function are also incorporated. At the National Agricultural Res e a rc h Center (NARC, http://ss.narc.go.jp/) at Tsukuba, Inoue applied a differential global positioning system (DGPS) and an optical fiber gyroscope (three axis) mounted on a 55-kW (75-HP) trac(a) (b) tor for tillage, and a Kalman filter Figure 17. (a) Total station and (b) navigation test in the field [31]. was used for estimation of the current position [31]. The accuracy of the DGPS was 0.15 m (sampling 90 speed: 1 Hz), and that of the optical 80 Fruit fiber gyroscope was 0.3°. A rotary tillage test was performed in the Leaf 70 field (100 × 160 m) at a speed of 1 Stem m/s. The offset error was within 0.1 60 Flower m, and that of a U-turn at the head50 land was 0.12 m. Nagasaka used a real-time kine40 matics GPS (RTKGPS) with an optical fiber gyro for an autonomous 30 rice planter [32]. As the GPS data 20 has a delay time (about 0.2 s) caused by communication with 10 the reference, compensation for this delay was added for real-time 0 300 500 700 900 1100 1300 1500 1700 1900 2100 2300 2500 position estimation. The GPS antenna was mounted on top of the Wavelength [nm] vehicle, resulting in an error of 0.1 Figure 18. Spectral reflectance of cucumber [40]. m at a roll angle of 3°. This inclination was also corrected. Steering Ehime University, a small automatic transport vehicle angle was determined according to the difference in attiequipped with a carriage self-correction mechanism was de- tude angle error. veloped for use in greenhouses [28]. At the National Grassland Research Institute (NGRI, In 1993, the Japanese Ministry of Agriculture, Forestry, http://ss.ngri.affrc.go.jp/), an autonomous tractor for forand Fisheries (MAFF, http://www.maff.go.jp) initiated the age production was developed using a fiber-optic gyroAgricultural Machine Development Project in which the scope and an ultrasonic Doppler speed sensor for position development of unmechanized machinery was promoted. identification [33]. Development of a tillage robot and driverless air blast Among manufacturers, the Crop Engineering System sprayer were started in BRAIN (Bio-oriented Technology Laboratory, Inc., was founded by Kubota Co., Ikegami Research Advancement Institute), and a driverless air Tsushinki Co., Ltd., and BRAIN. Application of a tracking lablast sprayer is now in field use [29]. Through this project, ser finder system and laser range sensor on automatic farm many technologies have been developed, and a total sta- machine systems has been performed for an autonomous tion (TOPCON Co. Ltd.) with automatic tracking for moving rice planter [34]. objects is now being used as the position sensor of a tillage This research is summarized in Table 5. Since the cost of robot in the field [30]. Fig. 17 shows the total station and RTKGPS is rapidly decreasing and the performance of image the navigation test in the field. The automatic tracking, po- processing is increasing, the combination of machine vision sition measurement, and data communication perfor- and RTKGPS appears to be the most promising system for the mances were sufficient at a distance of 500 m. The tillage future.

82


October 2001


Table 5. Results of navigation tests. Institute

Machine

Sensor

Velocity (m/s)

Errors (m)

References

University of Tokyo

Tractor

Vision

0.25

0.02

[24]-[26]

Hokkaido University

Tractor

Geomagnetic direction sensor

0.5

0.4

Kyoto University

Combine

Ultrasonic sensor

0.55

Ehime University

Transport vehicle

Self-carriage

0.5

Not described

[28]

MAFF (NARC)

Tractor

DGPS and optical fiber gyroscope

1.0

0.1

[31]

MAFF (NARC)

Rice planter

RTKDGPS and optical fiber gyroscope

0.8

0.15

[32]

MAFF (NGRI)

Tractor

Optical fiber gyroscope and ultrasonic Doppler speed sensor

1.2

1-2

[33]

BRAIN

Tractor

Image processing and laser range sensor

0.4

0.05

[30]

BRAIN

Speed sprayer

Guiding cable

0.7

0.1

[29]

Kubota

Mechanical weeding for rice

Laser range sensor

0.7

0.05

[34]

[27]

MAFF (Ministry of Agriculture, Forestry, and Fisheries) BRAIN (Bio-oriented Technology Research Advancement Institute)

Harvesting Robots Developments of harvesting robots were conducted in the United States, Europe, and Japan in the 1980s. In Japan, research on a harvesting robot for tomato was initiated in 1984 at Kyoto University, mainly by Fujiura and Ura, who received an award from the Japanese Society of Agricultural Machinery in 1991 [35], [36]. Since 1990, Okayama University has been leading the research in harvesting robots, such as for tomato, cucumber, grape, and strawberry crops [37]-[40]. In these robots, spectral reflectance was used for the discrimination of fruit from leaf and stem. Fig. 18 shows the spectral reflectance of cucumber fruit and leaf. The reflectance of the fruit is higher than that of leaf and stem in the near-infrared band; therefore, band-pass filters of 550 and 850 nm were used with a monochrome camera for the recognition of fruit. In this research, the cultivation types were also improved to discriminate the fruit from other parts. Fig. 19 shows a cucumber-harvesting robot; the stem of the cucumber was set inclined so that the cucumbers were separated from leaves and stems. A redundant manipulator is used for the harvesting to avoid obstacles such as stems or leaves. Robots for harvesting leaf vegetables such as cabbage are also being developed [41]. Little research in harvesting robots is currently under way in Europe and the United States because robotic perfor-

October 2001

mance is still low and the human operator is superior in cost and reliability. Therefore, even in Japan, continued interest will depend on new technological innovations.

Nursery Production Robots Nursery production robots such as transplanting and grafting robots are widely commercialized. In Japan, the grafted nursery, which has strong tolerance against injury by continuous cropping, is mainly used in the greenhouse. The ratio of each crop in the grafted nursery is 70% cucumber, 30% tomato, 50% eggplant, and 90% watermelon [42]. Although there is some variety in grafting methods, for the most part, the machines put together a scion and a rootstock using a clip, pin, and special bond adhesive method. Proper treatment after grafting is a necessity, typically requiring a dark chamber with high humidity; thus, an increase in the success rate is required, to over 90%. Grafting machines operate on the plants one by one or in one row at a time. The performance of the grafting machine is about 800-1,000 plants/hour, which is ten times that of human operators. A grafting robot is shown in Fig. 20. Plug-type seedlings are transplanted from a small tray, in which seedlings are planted at higher density, to a larger tray, and some plants are transplanted into separate pots. A transplanting machine is used for this operation whose per-



83

Figure 19. Cucumber harvesting robot [40].

Figure 20. Grafting robot (Komatsu).

formance is approximately 6,000 plants/hour [43]. Sensing of the stem is critical in this operation, and photo sensors or capacitance sensors are mainly used. Several types of these machines are already on the market. Although research in tissue culture robots was performed in the early 1990s in the United States, Europe, and Japan, the market for this machine was too small, and most of this research has been discontinued.

of image processing is increasing, a robotic system combining machine vision and RTKGPS appears to hold the most promise for the future. The nursery production robot, transplanting robot, and grafting robot are already in the marketplace, and many new technologies and innovations are being developed in this area.

Acknowledgment

Conclusions

The authors wish to express their appreciation to Prof. N. Sigrimis for the invitation to write this article.

In this article, we have discussed the application of intelligent approaches to optimization problems in agriculture in Japan. First, new algorithms, the PA and the LCA, derived from biosystems, were applied to search and optimization problems. Biosystems include numerous different natural phenomena, many of which are very peculiar and impressive. Many other algorithms are represented in biosystems that may prove useful in engineering applications; seeking useful engineering principles exemplified in biosystems is likely to be a fruitful path to advances in bioengineering. Second, an intelligent control system consisting of a decision system based on NNs and GAs and a feedback control system were applied to the optimization of the growth of hydroponic tomato plants during the seedling stage. The optimal four-step set points of nutrient concentration that maximize the TLL/SD ratio were successfully obtained using this decision system. The values of the TLL/SD ratio were 10-15% higher with the optimized control than with a conventional method. Good seedlings were obtained, with better reproductive growth potential. Finally, recent developments in intelligent agricultural robots in Japan were introduced. Although many robots for agricultural use are being developed or studied in Japan, most are far from ready for practical use. However, there is great demand for robotics and automation for crop management tasks, such as pesticide and herbicide application; thus, autonomous navigation in the field is a promising research area. As the cost of RTKGPS is rapidly decreasing and the performance

84

References [1] J.D. Farmer, N.H. Packard, and A.S. Perelson, “The immune system, adaptation, and machine learning,” Physica, vol. D22, pp. 187-204, 1986. [2] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Reading, MA: Addison Wesley, 1989. [3] J.B. Bowyer and R.C. Leegood, “Photosynthesis,” in Plant Biochemistry, P.M. Dey and J.B. Harborne, Eds. San Diego, CA: Academic, 1997, pp. 49-110. [4] H. Murase, S. Koyama, and A. Tani, “Finite element neural network for non-invasive plant water status monitoring,” Contr. Eng. Practice, vol. 2, no. 2, pp. 211-218, 1994. [5] M.Z.J. Gale and J. Ben-Asher, “Root aeration in a deep hydroponic system and its effect on growth and yield of tomato,” Scientia Horticulturae, vol. 19, pp. 213-220, 1983. [6] T. Fukuyama, T. Morimoto, and Y. Hashimoto, “Environment control of root system of melons cultivated in a computer controlled greenhouse,” Environment Contr. Biol., vol. 24, no. 1, pp. 9-20, 1986 (in Japanese with English summary). [7] Y. Hashimoto, “Recent strategies of optimal growth regulation by the speaking plant concept,” Acta Horticulturae, vol. 260, pp. 115-121, 1989. [8] Y. Hashimoto and K. Hatou, “Knowledge based computer integrated plant factory,” in Proc. 4th Int. Cong. Computer Technology in Agriculture, 1992, pp. 9-12. [9] Y. Hashimoto, “Applications of artificial neural networks and genetic algorithms to agricultural systems,” Comput. Electron. Agriculture, vol. 18, no. 2,3, pp. 71-72, 1997. [10] S. Chen, S.A. Billings, and P.M. Grant, “Non-linear system identification using neural network,” Int. J. Control, vol. 51, no. 6, pp. 1191-1214, 1990. [11] J.H. Holland, “Genetic algorithms,” Sci. Amer., pp. 44-50, July 1992. [12] D.L. Ehret and L.C. Ho, “Effects of osmotic potential in nutrient solution on diurnal growth of tomato fruit,” J. Exp. Botany, vol. 37, no. 182, pp. 1294-1302, 1986. [13] H. Luo and T. Kato, “Studies on the characteristics of seedlings raised in pots under various conditions and their productivity in eggplant and sweet pepper,” Environ. Contr. Biol., vol. 25, no. 3, pp. 91-96, 1987 (in Japanese).


October 2001


[14] T. Morimoto, T. Torii, and Y. Hashimoto, “Optimal control of physiological processes of plants in a green plant factory,” Contr. Eng. Practice, vol. 3, no. 4, pp. 505-511, 1995. [15] T. Morimoto, J. Suzuki, and Y. Hashimoto, “Optimization of a fuzzy controller for fruit storage using neural networks and genetic algorithms,” Eng. Applicat. Artif. Intell., vol. 10, no. 5, pp. 453-461, 1997. [16] T. Morimoto, W. Purwanto, J. Suzuki, and Y. Hashimoto, “Optimization of heat treatment for fruit during storage using neural networks and genetic algorithms,” Comput. Electron. Agriculture, vol. 19, pp. 87-101, 1997. [17] K.S. Narendra and K. Parthasarathy, “Identification and control of dynamical systems using neural networks,” IEEE Trans. Syst., Man, Cybernet., vol. 1, no. 1, pp. 4-27, 1990. [18] R. Isermann, S. Ernst, and O. Nelles, “Identification with dynamic neural,” in Preprints of 11th IFAC Symposium on System Identification, Fukuoka, Japan, 1997, vol. 3, pp. 997-1022. [19] T. Morimoto and Y. Hashimoto, “AI approaches to identification and control of total plant production systems,” Contr. Eng. Practice, vol. 8, no. 5, pp. 555-567, 2000. [20] D.E. Rumelhart, G.E. Hinton, and R.J. Williams, “Learning representation by back-propagation error,” Nature, vol. 323, no. 9, pp. 533-536, 1986. [21] T. Morimoto, W. Purwanto, J. Suzuki, and Y. Hashimoto, “Identification of cumulative fruit responses during the storage process using neural networks,” in Preprints of 11th IFAC Symposium on System Identification, Fukuoka, Japan, 1997, vol. 3, pp. 1555-1560. [22] R.G. Hurd, “The root and its environment in the nutrient film technique of water culture,” Acta Horticulturae, vol. 82, pp. 87-97, 1978. [23] Y. Mizrahi, E. Taleisnik, V. Kagan-Zur, Y. Zohar, R. Offenbach, E. Matan, and R. Golan, “A saline irrigation regime for improving tomato fruit quality without reducing yield,” J. Amer. Soc. Horticultural Sci., vol. 113, no. 2, pp. 202-205, 1986. [24] T. Torii, T. Kanuma, T. Okamoto, and O. Kitani, “Image analysis of crop row for agricultural mobile robot,” in Proc. AgEng96, Madrid, Spain, 1996, pp. 1045-1046. [25] T. Torii, A. Takamizawa, T. Okamoto, and K. Imou, “Vision-guided tractor,” in Proc. AgEng98, A-44, Oslo, Norway, 1998, pp. 1-8. [26] S. Noguchi, K. Ishii, and H. Terao, “Development of an agricultural mobile robot using a geomagnetic direction sensor and image sensors,” J. Agricultural Eng. Res., vol. 67, pp. 1-15, 1997. [27] M. Iida, M. Umeda, and M. Suguri, “Automated follow-up vehicle system for agriculture,” ASAE Paper 983112, Joseph, MI, pp. 1-9, 1998. [28] J. Yamashita, K. Satou, M. Hikita, T. Imoto, and T. Abe, “Development of an automatic guided vehicle for use in greenhouses and its traveling performance,” in Proc. 1st IFAC Workshop Mathematical and Control Applications in Agriculture and Horticulture, Matsuyama, Ehime, 1991, pp. 237-242. [29] K. Tosaki, S. Miyahara, T. Ichikawa, S. Taniai, Y. Mizukura, H. Moriki, and S. Miyashita, “Development of a microcomputer controlled driverless air blast sprayer,” in Proc. ARBIP95, Kobe, Hyogo, 1995, pp. 49-56. [30] O. Yukumoto, Y. Matsuo, and N. Noguchi, “Research on autonomous land vehicle for agriculture,” in Proc. ARBIP95, Kobe, Hyogo, 1995, vol. 1, pp. 41-48. [31] K. Inoue, K. Otsuka, M. Sugimoto, and N. Murakami, “Estimation of place of tractor and adaptive control method of autonomous tractor using INS and GPS,” in Proc. BIO-ROBOTICS 97, Gandia, Valencia, 1997, pp. 21-24. [32] Y. Nagasaka, K. Taniwaki, R. Otani, K. Shigeta, and Y. Sasaki, “The Development of autonomous rice transplanter,” J. Jpn. Soc. Agricultural Machinery, vol. 61, no. 6, pp. 179-186, 1999. [33] A. Okado, M. Ishida, K. Imou, H. Takenaga, and N. Itokawa, “Development of an autonomous tractor for forage production,” in Proc. 3rd Int. Symp. Artificial Life and Robotics, Beppu, Oita, 1998, pp. 238-241. [34] J. Yoshida, “A study on the automatic farm machine system for rice,” Osaka, Crop Engineering System Laboratory, Inc., 1997. [35] N. Kawamura, K. Namikawa, T. Fujiura, and M. Ura, “Study on agricultural robot,” J. Jpn. Soc. Agricultural Mach., vol. 46, no. 3, pp. 353-358, 1984. [36] T. Fujiura, M. Ura, N. Kawamura, and K. Namikawa, “Fruit harvesting robot for orchard,” J. Jpn. Soc. Agricultural Mach., vol. 52, no. 2, pp. 35-42, 1990.

October 2001

[37] N. Kondo, M Monta, Y. Shibano, and K. Mohri, “Two finger harvesting hand with absorptive pad based on physical properties of tomato,” Environ. Contr. Biol., vol. 31, no. 2, pp. 87-92, 1993. [38] N. Kondo, Y. Shibano, K. Mohri, M. Monta, H. Nakamura, and S. Arima, “Basic studies on robot to work in vineyard,” J. Jpn. Soc. Agricultural Mach., vol. 55, no. 6, pp. 85-94, 1993. [39] N. Kondo, Y. Shibano, K. Mohri, M. Monta, and S. Arima, “Studies on visual sensor for cucumber fruit detection,” Environ. Contr. Biol., vol. 31, no. 2, pp. 93-100, 1993. [40] N. Kondo and K.C. Ting, Eds. Robotics for Bioproduction Systems, Joseph, MI: ASAE, 1998. [41] N. Murakami, K. Otsuka, K. Inoue, and M. Sugimoto, “Development of robotic cabbage harvester,” J. Jpn. Soc. Agricultural Machi., vol. 61, no. 5, pp. 85-92, 1999. [42] T. Tomita and H. Wadake, “Grafting robot for vegetable seedling,” J. Soc. Instrum. Contr. Eng., vol. 37, no. 2, pp. 99-100,1998. [43] H. Yamada, “Development of transplanting robot,” J. Roboti. Mechatronics, vol. 11, no. 3, pp. 227-230, 1999.

Yasushi Hashimoto received the Ph.D. in agricultural engineering from the University of Tokyo in 1967. Since 1985, he has been a Professor of Biomechanical Systems, Ehime University. He is currently engaged in research and teaching in mechanization, automation, and informatics in plant production systems. He is President of the Japanese Society of Environment Control in Biology and Chairman of the Coordinating Committee on Life Support Systems of IFAC. Haruhiko Murase received his Ph.D. in agricultural engineering from Michigan State University in1977. After receiving his Ph.D., he joined the faculty of Osaka Prefecture University, Japan, where he is now a Professor of Agricultural Engineering. Tetsuo Morimoto received the Ph.D. degree in agricultural engineering from Ehime University, Japan, in 1993. From 1977 until 1994, he was an Assistant Professor and since 1994 an Associate Professor of Biomechanical Systems, Faculty of Agriculture, Ehime University. His research interests include the identification and optimal control of plant production processes using intelligent approaches such as neural networks and genetic algorithms. Toru Torii received the B.A. degree in agricultural engineering in 1980, the M.S. degree in mechanical engineering for production in 1982, and the Ph.D. degree in 1992 from the University of Tokyo. From 1982 to 1984, he was with Mitsubishi Motors Co., and from 1989 to 1999, he was an Assistant Professor in agricultural engineering at the University of Tokyo. Since 1999, he has been an Associate Professor in precision engineering at the University of Tokyo. His current interests are microfluidic devices and micromechatronics devices for lab on a chip.



85