FSPM04
POSTERS - SESSION 3
Modeling the soybean growth in different amount of nitrogen, phosphorus and potassium using neural network A. Suratanee, S. Siripant, C. Lursinsap Advanced Virtual and Intelligent Computing (AVIC) Research Center, Department of Mathematics, Chulalongkorn University, Phayathai Road, Bangkok, 10330, Thailand.
[email protected],
[email protected],
[email protected]
This paper proposed a simulation model of soybean growth which is effected by major nutrient factors, nitrogen, phosphorus and potassium. A feedforward neural network is used as a basis of the modelling. The combination of different percentage of nitrogen, phosphorus, potassium, time steps and the collected height data of the soybean are used as inputs. The model can predict the height at designated time intervals, whereby the result can be visualized with L-systems. Keywords: Simulation and Visualization, Soybean Growth, Neural network modeling, Solution culture, L-system. Introduction The simulation of plant developments has been published in the context of computer graphics, treating a plant as a closed system without considering the interaction of plant and its environment [6]. The interaction between a plant and its environment factors has been proposed in a specific assumption for computer graphics [4]. Most of the previous works, regarding virtual plants, relate the diversity of the plant growth caused by the environment factors (temperature, humidity, photoperiod [7, 8], amount of light [1], etc.), whereas the effect of nutrient on the growth rate in terms of mathematical equations is discussed in this research. Solution culture or hydroponics, growing plants in a defined nutrient solution, is the principal experimental system for studying plant nutrient requirements. Over the years a large number of nutrient solutions have been formulated for studying the nutritional requirements of plants. Most modern formulations are based on a solution originally developed by D.R. Hoagland, a pioneer in the study of plant mineral nutrition. Individual investigators may introduce minor modifications to the composition of the nutrient solution in order to accommodate specific needs. Such formulations are commonly refered to as modified Hoagland's solution [3]. Neural networks are similar to nonlinear regression, but they are much more robust and can expose hidden relationships in large bodies of information by using pattern recognition theory. They have successfully been used in biological applications to predict processes such as optimum temperatures for greenhouses, insect pest treatment thresholds, recognition of patterns from digital images [5] and predicts pH and electrical conductivity (EC) changes in the root zone of lettuce [2]. The objective of this work was to develop and validate a neural network model to be able to predict the height of soybean which grown in the nutrient solution of hydroponics. The result of this model will provide a first step toward solving the further problem of using such information to develop a model of nutrient requirements for other plants. The rest of the paper is organized into three sections. Section 2 discuss how the data are collected and modeled by a neural network. Section 3 gives the experimental results. Section 4 concludes the paper. Material and methods Experiment Methods The solution culture or hydroponics for growing soybean ( Glycine max (L.) Merr. ) and the data were collected from soybean growth. In this work, we are interested in determing the primary nutrients requirements of plants. There are numerous nutrient solution formulations described in scientific papers on plant nutrition, books and articles on hydroponics [3]. Some are designed for general use (like Hoagland's solution), and others for specific plants. These formulas are not suitable
4th International Workshop on Functional-Structural Plant Models, 7-11 june 2004 –Montpellier, France Edited by C. Godin et al., pp. 130-133
Modeling the soybean growth using neural network
131
for our study i.e. soybean. Hence, the mixture of nutrient formulas must be specifically prepared in our laboratory for this study [3]. The complete nutrient solution contains all essential minerals for plant growth, while trace elements are provided by impurities in the chemicals used. All solutions are made up of distilled water. In our experiment, we made a nutrient solutions which are deficient nitrogen, phosphorus or potassium from complete nutrient solution, by varying amount of deficient of nitrogen, phosphorus and potassium as 0, 50, 100 for each nutrient to the complete solution N-P-K. The assigned of complete nutrient solution is 100-100-100, which mean 100 percent of nitrogen, 100 percent of phosphorus and 100 percent of potassium, respectively. Thus 100-100-50 mixture means that complete nutrient solution has less potassium 50 percent. All of percentage of N-P-K deficient formulas are shown in Table 1. Network Design and Data Sets Data sets used to train and test a neural network ( NN ) are collected from an actual soybean. These data concern the internode length corresponding to the time of its life cycle. The actual data, which constituted inputs to the NN model, are obtained daily for 66 days. The NN model is used as the basis of modeling. The network has five inputs, namely, (1) percentage of nitrogen-deficient (2) percentage of phosphorus-deficient (3) percentage of potassiumdeficient (4) time step(day) (5) length of internode, and one output as the length of internode in next step time. After the training process, its performance and generalization capabilities were evaluated. Nitrogen
Table 1. Formulas of Nitrogen, Phosphorus and Potassium deficiencies 100 50 0 100 50 0 100
50
0
Phosphorus
100
100
100
50
50
50
0
0
0
Potassium
100
100
100
100
100
100
100
100
100
Nitrogen
100
50
0
100
50
0
100
50
0
Phosphorus
100
100
100
50
50
50
0
0
0
Potassium
50
50
50
50
50
50
50
50
50
Nitrogen
100
50
0
100
50
0
100
50
0
Phosphorus
100
100
100
50
50
50
0
0
0
0
0
0
0
0
0
0
0
0
Potassium
Results The model simulation In preliminary result, Figure 1 and Figure 3, are graphs of the measured height data and values predicted by the NN model in 100-100-0 and 50-100-0 formula, respectively, are presented. This shows that the neural network model gives a good prediction to the actual data. The error of this prediction ,100-100-0 and 50-100-0, expressed as a percentage error, are 0.87% and 0.52%, respectively. Visualization The development of soybean in different amount of nitrogen, phosphorus and potassium is visually presented. The L-systems qualitative model represents the plant topology and development. We evaluated the soybean growth and made some parameters adjustment such as the size of each internode, to render a more realistic look of plant growth. Figure 2, 4 shows the development of soybean growth at 66 days by the sum of internode length (height) are given from prediction values in each case of nutrient formula.
POSTERS - SESSION 3
132
A. Suratanee et al.
Conclusions A predictive method that uses a supervised feedforward neural network to model the height of soybean growth in solution culture was developed. More specifically, the artificial neural network was applied successfully in a model that predicts the height of the soybean. In addition, this work builds a link between artificial intelligence, hydroponics systems and computer graphics. With its encouraging results, it also opens the way to further development and investigation of “intelligent” systems in the field of agriculture, and computer graphics, which will lead to more precise and productive cultivation in agriculture systems.
Figure 1. Comparison of real measurements and neural network predictions in 100-100-0 Formula
Figure 2. Visualization of soybean growth in 100-100-0 at 66 days
Figure 3. Comparison of real measurements and neural network predictions in 50-100-0 Formula
Figure 4. Visualization of soybean growth in 50-100-0 at 66 days
Acknowledgement This work is partially supported by National Electronics and Computer Technology Center (NECTEC) of Thailand.
References Bedrich Benes, “An Efficient Estimation of Light is Simulation of Plant Development”, Dissertation thesis, Faculty of Electrical Engineering, Czech Technical University, Prague, 1998. K.P. Ferentinos and L.D, “Albright. Predictive neural network modeling of pH and electrical conductivity in deep-trough hydroponics”. Trans. ASAE 45(6): 2007-2015. 2002. N. Angkinand and S. Chadchawan, “Plant physiology laboratory”, Department of Botany, Chulalongkorn University, Bangkok, 2000.
SESSION 3 – POSTERS
Modeling the soybean growth using neural network
133
N. Chiba, K. Ohshida, K. Muroaka, and S. Nobuji. “A Growth Model Having the Abilities of GrowthRegulations for Simulating Visual Nature of Botanical Trees”, Computer & Graphics, 18:469479, 1994. Robert M. Peart and R. Bruce Curry, “Agricultural systems modeling and simulation”, Marcel Dekker, Inc., New York, 1998. S. Chuai-Aree, S. Siripant and C. Lursinsap, “Animation plant growth in L-system by parametric functional symbols”, Proceeding of International Conference on Intelligent Technology 2000, December 2000, pages143-135 . Sikora S., Steinberg D. and Lattaud C., Integration of simulation tools in on-line virtual worlds, in Proceedings of the 2nd International Conference on Virtual Worlds, Jul. 2000, pages 32-43. Steinberg D., Sikora S., Lattaud C., Fournier C. and Andrieu B., Plant growth simulation in virtual worlds : towards online artificial ecosystems, in Proceedings of the 1st Workshop on Artificial Life Integration in Virtual Environments, Lattaud C. ed., Sep. 1999, pages 19-25.
POSTERS - SESSION 3
