As part of my Intelligent Systems course in my Masters program, I am working on a project that joins two intelligent systems together. I won’t post many of the finer details of the project (check out my conference paper once I get around to writing it), but it is a rule-based and genetic algorithm hybrid.

The idea here is not create a “super-agent” that can win at games, but more to simulate personality in NPC characters. I’ve yet to generate a decent set of experiment results, but things are looking good so far. I have my test software written and I’m about to add my genetic algorithm library to the project to see what happens.

This is a model of environmental control strategies based on plant responses using intelligent machine vision technique.

Neural network was used to identify the relationship between plant moisture content and image features. Sunagoke moss water content prediction using relevant feature selection can also be applied as environmental control based on plant response. The figure below shows the neural network control system with a bio-feed back loop containing another neural network which converts the pictorial information into moss water content. The vegetative system is the controlled system which controlled variable is the moss water content. The control element of this system is the environment surrounding the plant with which the control inputs are temperature, humidity, CO2 concentration, water and light intensity.

Substantial increases in yield could be possible if irrigation water was applied at the most appropriate time and accurate amount of water to prevent excessive drought stress.

With the increase in the cost of energy required to pump and move water to desired locations, coupled with the decrease of available water for irrigation, it is essential to attain the maximum benefit from each quantity of water used for irrigation. Plant drought stress detection is thus of great importance.

Efficient control of water delivery to crops optimizes production quality while minimizing inputs to the system. Over-watering crops not only wastes water, it also leads to higher expenses in pumping energy, heat energy to dry the environment, and disease and pest treatments required due to damp conditions. Accurate water management in controlled-environment crop production is accomplished by irrigation when necessary and in the correct quantity.

Human visual assessment of crop stress is qualitative at best, with the terms “good” or “poor” frequently used to describe crop condition. To improve the monitoring and irrigation process, automatic monitoring and quantitative describing of plant water content are necessary. Computer-controlled irrigation systems can utilize a range of data to achieve accurate delivery of water according to crop requirements. Automated control systems that use timers, supply-based measurements of water or demand-based models of expected water use do not attempt to measure the actual water content of the plants they serve. A more accurate control system would use the plant status as feedback for controlling the growth environment and irrigation. Hashimoto et al (1985) developed measurement and instrumentation in greenhouse to investigate ways to use the crop as a feedback for greenhouse control systems with the “speaking plant” approach.

At present, because of the advances in electronic technology, machine vision can be applied to research on the development of an automatic prediction. Machine vision enables to handle a large amount of raw data and perform remote prediction. On the other hand, there are conventional image recognition methods such as statistical method and logical method of computation. However, they are not suitable for determining the water content in Sunagoke moss because of the complexity and ambiguity involved.

Sunagoke moss is a biological product. Although Sunagoke moss may be cultivated in the same way, their shape, size, color and other characteristics are different, therefore it has various patterns in ET. Using machine vision to evaluate the water content of Sunagoke moss and extraction of image features become an important element because the water content of Sunagoke moss does not follow quantitative rules. It is important that a machine vision prediction system has learning process functions. Such a system enables to overcome the weakness of traditional type systems. Therefore, the researchers have developed an advanced prediction system using machine vision, involving image processing, image feature extraction, neural network and computer technologies.

The model study as shown in figure 1 including: analyzing water content using weight measurement, image acquisition using digital camera, image processing and image analysis, feature extraction to extract color, morphology and textural features and the last is feature selection. The outputs of feature selection are selected subset of features and neural network (NN) weight that can be applied for neural network based irrigation.

Fig. 1 Model study of neural-network-based irrigation system

If the relevant features have already selected, and optimum NN-weights have been determined, the flowchart of the prediction system software is shown in figure 2. The system can estimate water content according to their image features using the learned standard pattern. This software displays learning and prediction components as follows:

Fig. 2 Flowchart of machine vision prediction system software

The proposed irrigation control design of Sunagoke moss is illustrated in figure 3. Pictorial information from Sunagoke moss (selected features from Feature Selection) can be utilized as the input of Back-propagation Neural Network (BPNN) to detect dry area and the output is the amount lack of water. The water sprayer will irrigate the dry part of Sunagoke moss in accurate places and accurate amount of water. The decrement of dry area and optimum water condition can bring Sunagoke moss into optimum photosynthesis process. The main objective of this precision irrigation system is to optimize the water supply of Sunagoke moss and stabilize the amount of water content between 2 gg^-1 and 2.5 gg^-1.

Fig. 3 Proposed irrigation control design for Sunagoke moss

Water content has relation with optimum photosynthesis. Therefore, if we can determine part of area which is lack of water through positioning (coordinate system), then we can automatically give appropriate water at that precise area, so the decrement of dry area can bring Sunagoke moss into optimum photosynthesis process. Figure 4 shows the structure/design of NN-based irrigation system for Sunagoke moss which will be applied for Sunagoke moss cultivation in tunnel.

Fig. 4 General design of NN-based Irrigation system for Sunagoke Moss

In Genetic-neural algorithm methods, a search procedure in the space of possible feature subsets is defined, and various subsets of features are generated and evaluated.

In wrapper methods, the evaluation of a specific subset of features is obtained by training and testing a specific classification/prediction model, rendering this approach tailored to a specific classification/prediction algorithm. To search the space of all feature subsets, a search algorithm is then wrapped around the classification/prediction model. However, as the space of feature subsets grows exponentially with the number of features, heuristic search methods are used to guide the search for an optimal subset. These search methods can be divided in two classes: deterministic and randomized search algorithms. In the wrapper approach proposed by Kohavi and John (1997), the feature subset selection algorithm exists as a wrapper around the induction algorithm. The feature subset selection algorithm conducts a search for a good subset using the induction algorithm itself as part of the function evaluating feature subsets. The idea behind the wrapper approach, shown in figure 1, is simple: the induction algorithm is considered as a black box. The induction algorithm is run on the data-set, usually partitioned into internal training and holdout sets, with different sets of features removed from the data. The feature subset with the highest estimated value is chosen as the final set on which to run the induction algorithm. The resulting classifier is then evaluated on an independent test set that was not used during the search.

Fig.1 The wrapper approach to feature subset selection. The induction algorithm is used as a black box by the subset selection algorithm.

Advantages of wrapper approaches include the interaction between feature subset search and model selection, and the ability to take into account feature dependencies. A common disadvantage of these techniques is that they have a higher risk of over fitting than filter techniques and are very computationally intensive.
Genetic Algorithms (GAs) are search algorithms based on the mechanics of natural selection and natural genetics (Goldberg, 1997). GA are different from more normal optimization and search procedures in four ways: (1) GAs work with a coding of the parameter set, not the parameters themselves; (2) GAs search from a population of points, not a single point; (3) GAs use payoff (objective function) information, not derivatives or other auxiliary knowledge; and (4) GAs use probabilistic transition rules, not deterministic rules.
Genetic algorithms are adaptive search techniques based on the principles of natural selection in biology. The GAs are stochastic global search methods that mimic the metaphor of natural biological evolution. They employ a population of competing solutions evolved over time to converge to an optimal solution. Effectively, the solution space is searched in parallel, which helps in avoiding local optima. For feature selection, a solution is typically a fixed length binary string representing a feature subset the value of each position in the string represents the presence or absence of a particular feature. The algorithm is an iterative process where each successive generation is produced by applying genetic operators such as crossover and mutation to the members of the current generation. Mutation changes some of the values (thus adding or deleting features) in a subset randomly. Crossover combines different features from a pair of subsets into a new subset. The application of genetic operators to population members is determined by their fitness (how good a feature subset is with respect to an evaluation strategy). Better feature subsets have a greater chance of being selected to form a new subset through crossover or mutation. In this manner, good subsets are “evolved” over time.

A typical series of operations carried out when implementing a GAs paradigm is:
• Initialize the population;
• Calculate fitness for each chromosome in population;
• Reproduce selected chromosomes to form a new population;
• Perform crossover and mutation on the population;
• Loop to second step until some condition is met.

Initialization of the population is commonly done by seeding the population with random values. The fitness value is proportional to the performance measurement of the function being optimized. The calculation of fitness values is conceptually simple. It can, however, be quite complex to implement in a way that optimizes the efficiency of the GAs search of the problem space. It is this fitness that guides the search of the problem space. After fitness calculation, the next step is reproduction. Reproduction comprises forming a new population, usually with the same total number of chromosomes, by selecting from members of the current population using a stochastic process that is weighted by each of their fitness values. The higher the fitness, the more likely it is that the chromosome will be selected for the new generation. One commonly used way is a roulette wheel procedure that assigns a portion of a roulette wheel to each population member where the size of the portion is proportional to the fitness value. This procedure is often combined with the elitist strategy, which ensures that the chromosome with the highest fitness is always copied into the next generation. The next operation is called crossover. To many evolutionary computation practitioners, crossover is what distinguishes GAs from other evolutionary computation paradigms. Crossover is the process of exchanging portions of the strings of two ‘‘parent” chromosomes. An overall probability is assigned to the crossover process, which is the probability that given two parents, the crossover process will occur. This probability is often in the range of 0.65-0.80. The final operation in the typical GAs procedure is mutation. Mutation consists of changing an element’s value at random, often with a constant probability for each element in the population. The probability of mutation can vary widely according to the application and the preference of the person exercising the GAs. However, values of between 0.001 and 0.01 are not unusual for mutation probability.

In this research, a combined genetic-neural algorithm was developed for FS based on the neural network pattern recognition. We have extracted 50 features including: color (RGB and Grey level), morphological features (green area, perimeter and browning index) and textural features. Each individual in the population represents a candidate solution to the feature subset selection problem. There were (2⁵⁰-1) possible feature subsets. The values of the  Back-propagation Neural Network (BPNN) inputs were the normalized image features that are between 0 and 1. One hidden layer (20 nodes) was used in BPNN. One output of BPNN was used to determine Sunagoke moss water content. The number of data have been used in BPNN were 124 image data with 10% data as testing set. The selected features are the inputs of the BPNN, which are used for water content prediction. In GAs for Feature Selection (FS) involves many generations. In each generation, evaluation of an individual (a feature subset) involves training neural networks. In this step, a binary vector of dimension 50 represents the individual in the population. The chromosome define contains 50 genes, one gene for each feature, which can take 2 values. A value of 0 indicates that the corresponding feature is not selected, while a value 1 means that the feature is selected. An initial population of chromosomes is randomly generated. Crossover was performed by 2 points real value crossover. Two points were selected randomly. Mutation process was also conducted randomly. Some best chromosomes were kept to be used in the next generation. Looping process was done until the fitness function converged to find best feature subset. The roulette wheel selection strategy was also used in the algorithm for FS. The parameter settings were: population size: 50; probability of best chromosomes: 0.2; probability of crossover: 0.8; probability of mutation: 0.01 and number of generation was developed until the fitness function converged. The fitness of the chromosome was calculated according to the prediction rate of the selected subset of features for predicting Sunagoke moss water content, as it is shown in figure 2.

Fig.2 Feature selection using combined genetic-neural algorithm.

Comparison between Filter and Wrapper Methods

We have investigated both filter and wrapper methods for feature selection. The research was conducted by the prediction rate based on testing set to calculate the fitness for reproduction of GAs. The number of hidden layer (1 hidden layer) and hidden nodes (20 nodes) were determined to be effectively used in BPNN structures to achieve the best prediction of water content. The results of FS are described as following. First of all, in most cases, the accuracy of prediction performance using five proposed FS method were greatly improved. It was shown in table 1 that there were some improvements in case of number of features used and the prediction error between methods using FS and method without FS.

Among the filter methods, Chi-Squared (X²) method got the highest performance for prediction with the validation error 0.009 and 31 features selected followed by Mutual Information (MI) with validation error 0.018 and 35 features selected. Corelation-based Feature Selection (CFS) and Linear Regression (LR) reached the minimum validation error approximately 0.023 and 0.020 and number of features selected 10 and 30 respectively.

Overall the results, the experiments which were conducted using filter method, the results were much worse than using wrapper method (genetic-neural algorithm). Genetic-neural algorithm performed better compare to filter method with reliable improvement value of 76.6%. When analyzing the results of the genetic-neural algorithm, it is noticed that the lowest prediction testing error (MSE: 0.0021) was achieved by BPNN with a few sets of selected features (27 features). GAs shows that the fitness function converged after 120 generations using 50 populations in each generation. Fig. 4 shows the frequency selected for each relevant feature in every generation. There were nine most selected features with the frequency selected rate above 90%. The most contributed features to predict water content were: grey energy with frequency selected 100%, green/red ratio variance (98.2%), blue inverse difference moment (97.1%), red correlation (96.9%), red mean value (95.7%), red cluster tendency (93%), blue homogeneity (91.2%), green entropy (91.1%) and grey variance (90.6%). Genetic-neural algorithm has successfully selected relevant image features to determine water content of Sunagoke moss.

Table 1. Comparison analysis of five feature selection methods.

Fig. 3 Fitness functions of genetic algorithm for feature selection.

Fig. 4 The selection rate of every selected feature.

If you want to know more about How this genetic-neural algorithm works, you can download PPT file HERE.

Providence is one of the oldest cities in America and it has the architecture to prove. It has a larger percentage of buildings on the National Historic Register than any other U.S. city, with scores of immaculately preserved Colonial, Federal, Greek Revival and Victorian homes and buildings.

Story was begun when I arrived at Detroit Wayne County Airport Michigan, as usual, maybe because of “my name”, I should go to interrogation room again…(once more Again!!), and the immigration officer started to take my promise that I would do nothing bad in “their” country….what a nice guy and polite officer, but let’s forget it. It’s only a matter of “culture”. But this time I have prepared by arranging my flight schedule to Providence a little bit longer because I already know that I will be in Interrogation room again.

Detroit Wayne County Airport Michigan

The ASABE conference was held in Rhode Island Convention Center starting form June 28th until July 2nd.

Rhode Island Convention Center

The Techincal Session including:

1. Pesticide Application technologies
2. Biomass-Harvesting, storage, transporting, economic considerations and producers sustainability
3. Advances in soil dynamics
4. Development and performance of yield monitors
5. Soil erosion
6. Advances in soil and water engineering
7. Advance of numerical methods in food and process engineers
8. Wireless networks and applications
9. Bioimaging and machine vision for food safety, quality and processing
10. Air pollutant measurement and control from animal operations
11. Innovative energy strategies for controlled environments
12. Agricultural waste management
13. Ergonomics safety and health division
14. Sensing and Automation in food processing
15. Ethanol and biodiesel industry process improvement
16. biomass feedstock engineering issues
17. Emerging technologies in food processing
18. Physical properties of agricultural materials in relation to processing
19. Market-based approach to ecosystem management
20. Innovations in undergraduate and graduate education
21. Engineering technology and standards for developing countries
22. New development in cultural practices
23. Management tools for agricultural systems
24. Automated guidance of agricultural vehicles
25. Statistical techniques in hydrology
26. streambank erosion-processes and modeling
27. Soil erosion research
28. Soil erosion on non-agricultural lands
29. Advances in remote sensing applications for ET estimation and mapping
30. Advances in instrumentation and control
31. Advances in machine vision techniques
32. Structures and environment
33. Emerging technologies for agricultural safety and health
34. Food and process engineering
35. Emerging fermentation and purification technologies
36. Challenges in lignocellulosic conversion
37. Biological engineering
38. Forest based biomass: harvesting, processing, transport and storage
39. Power and machinery
40. Climate forecast for water resources decision support
41. modeling and evaluation of conservation practice
42. Precision irrigation
43. Micro/subsurface drip irrigation
44. Information and electrical technologies
45. Bulk solids and biomass storage systems
46. Animal responses to environmental factors
47. Greenhouse and nursery technology applications
48. Air Quality modeling
49. Nitrogen and phosphorus management for manure and litter
50. Food and safety engineering
51. Biofuel thermal-process development research
52. Drying and processing crops
53. Advances in multispectral and hyperspectral properties in food and agricultural products
54. Ethanol development research
55. Physical and chemical properties of biresource products and co-products
56. Bioenergy
57. Pesticide application technologies
58. State of the art on-the-go soil strength measurement
59. Spectral sensing for precision agriculture
60. Geospatial technologies for management decision making
61. Objective evaluation of hydrologic model
62. Watershed-scale nutrient cycling and sediment transport
63. Drainage research
64. Advances in irrigation
65. Sensors for automation and biorobotics
66. Air quality as related to agricultural waste management
67. Agri-industrial facility design and operation
68. Environment control for animal facilities
69. Advances and applications of nondestructive testing of agricultural and food products
70. Microwave and radio frequency heating in agricultural and food processing
71. Handling and storage of coproducts for energy and grains
72. Application of ecological engineering
73. Agricultural system modeling
74. Advances in cotton engineering
75. Fruit and vegetable engineering
76. Drainage and water quality
77. Irrigation management
78. Hydrology and biogeochemical process and modeling challenges in forest ecosystems
79. Hyperspectral and multispectral imaging
80. Heat and power generation from solar, wind, geothermal and other renewable energy systems
81. Water management and manure treatment on dairies
82. Innovative technologies for organic farming
83. Environmental bioconversion processes
84. Variable rate technology for precision agriculture
85. Microbial pathogen fate and transport
86. Drainage modeling at the field and watershed scale
87. Vegetative treatment systems
88. Ergonomics and human factors
89. Advances in coproduct and byproduct management and utilization
90. Hydrogen and biofuel development research
91. Wetland engineering

This time I presented my paper in Precision Irrigation session.

Oral Presentation

Poster session

Intelligent Irrigation Control Using Color, Morphological and Textural Features in Sunagoke Moss

A non-invasive sensing technique for monitoring Sunagoke moss water conditions was proposed. This paper describes the design and development of precision irrigation control method by incorporating color, morphology and RGB color co-occurrence matrix (CCM) textural features. The objective of this study was to develop a model of artificial neural network and made comparison analysis of the color, morphology and textural features to determine appropriate combination of pictorial features to accurately predict water content. Optimum condition of Sunagoke moss based on photosynthesis rate, color features, morphological features and textural features can be achieved between 2 g/g- 2.5 g/g water content. Neural network model performance was tested successfully to describe the relationship between water content and image features (color, morphology and textural features). This system is helpful to explore the new way of water spraying in moss plant factories based on computer vision. It proposes the water irrigation technology of the plant factory to realize the automation and precision farming. Precision water and nutrition spraying system based on computer vision is very important, not only for spraying the water and nutrition scientifically, but also for improving the efficiency of spraying and decreasing the non- or off-target of moss to prevent from over watering.
Keywords: Image analysis, artificial neural network, irrigation system, control, plant factories, precision farming.

Full paper can be downloaded HERE.

A little bit about my journey to Boston:

Boston Station

Massachusetts Institute of Technology

Harvard University

]]> http://samuraielit.wordpress.com/2008/11/27/17th-ifac-world-congress-seoul-korea-2008/ Thu, 27 Nov 2008 11:44:17 +0000 samuraielit http://samuraielit.wordpress.com/2008/11/27/17th-ifac-world-congress-seoul-korea-2008/

The IFAC World Congress held in Seoul, Korea in 2008 is the 17th of the triennial IFAC World Congress series….a record number of more than 3700 papers have been submitted and approximately 2700 have been accepted. All of these papers was presented in 397 oral sessions and 7 poster sessions over the five day period, all in one conference venue of COEX, Seoul, Korea.

Seoul is a city of infinite discoveries. With a history of 600 years, Seoul is a city, where the 5000-year Korean traditional and the modern technologies co-exists in a perfect harmony.

Incheon Airport, Seoul

Incheon International Airport (ICN), the ‘Winged City’, is an airport development located on reclaimed land approximately 32 miles from downtown Seoul, South Korea.

Deoksugung Palace

The main houses of the former Deoksugung Palace were Prince Wolsandaegun’s residence who was King Seongjong’s elder brother. About a century later, when King Seonjo returned to Seoul in 1593 after fleeing the Japanese invasion of 1592, he established a temporary residence here. Haenggung in Jeongreung-dong, was renamed Gyeongungung In 1611, by the next ruler, Gwanghaegun. After Gwanghaegun moved to Changdeokgung Palace in 1615, the palace remained vacant for about 200 years.
In 1897, King Gojong returned to power, after having spent two years at the Russian legation in the wake of the assassination of his queen, Empress Myeongseong. He took up residence at Changdeokgung Palace, establishing an independent ‘Daehanjeguk Empire’.
From that time until 1907, it was the center of historical turbulence. In 1907, King Gojong was forced to abdicate his throne, although he continued to live at the palace until he passed away in 1919. Since 1907, the palace has been known as Deoksugung.

Seoul Tower

Viewable from almost anywhere in Seoul, Seoul Tower serves as an excellent landmark. Built on a 262 meter peak in Namsan Park, the tower reaches to 480 meters above sea level. When the weather and pollution levels cooperate, visiting the observation tower (370 meters above sea level) allows you to view the entire city and surrounding areas.

Korean Cuisines

The venue for the 17th World Congress is COEX Convention Center which is located within the World Trade Center Seoul Complex in the central business area of Seoul, Korea.

COEX Convention Center

Robot Demonstrations

MAHRU of KIST

MAHRU is the world’s first network-based humanoid. It has been endowed with artificial intelligence through a network. Unlike other famous humanoids, such as ASIMO, MAHRU focuses on network-based intelligence by using the network infrastructure, where Korea has world-class strengths.

POMI of ETRI

POMI (Penguin Robot for Multimodal Interaction) is a five-senses-mounted emotional expression robot that has been endowed with artificial intelligence and active reaction modules. POMI was developed by the U-Robot Research Division of ETRI and the hardware platform company.

EveR-1 of KITECH

The performance of the android, EveR-1, developed in KITECH (Korea Institute of Industrial Technology) was demonstrated. The appearance of EveR-1 is based on Korean Female. EveR-1 is not bipedal, but is capable of motion from her torso up.

Robot Soccer Demonstration Game of FIRA

In this Conference, I presented my paper in Greenhouses and controlled agricultural production session.

Water Irrigation Control for Sunagoke Moss Using Intelligent Image Analysis

A novel technique suitable for noninvasive measurements of moss water content is presented. In this paper, color image sensing is applied for measuring moss water content. Sunagoke moss Rhacomitrium canescens has been utilized as an active greening material to mitigate the urban heat island effect. The goal of this paper is to develop an intelligent image analysis system for water irrigation optimal control in Sunagoke moss. The combination of RGB components (green/red ratio, blue index, blue value and green index) using statistical pattern recognition can estimate water content and define the distribution of water condition in every pixel of Sunagoke moss images. The combination of colour image sensing and Artificial Neural Network (ANN) successfully described the relationship between water content and colour features i.e. average green index, average blue index, blue mean value, browning area index, green canopy index and average green/red ratio. This system is helpful to explore a new way of water spraying in Sunagoke moss plant factories based on computer vision. We propose a water irrigation technology of plant factory to realize the automation and precision farming. Precision water spraying system based on computer vision is important, not only for spraying the water scientifically, but also for improving the efficiency of spraying and decreasing the non- or off-target spraying to prevent over watering.
Keywords: Image analysis, artificial neural network, optimal control, plant factories, precision farming.

Full paper can be downloaded HERE.

]]> http://samuraielit.wordpress.com/2008/11/27/jsabees-matsuyama-ehime-university-2008/ Thu, 27 Nov 2008 04:57:36 +0000 samuraielit http://samuraielit.wordpress.com/2008/11/27/jsabees-matsuyama-ehime-university-2008/

….it was a JSABEES Annual Conference held in Matsuyama (Ehime University) in September 2008.

The Conference focused on Basic and Applied Sciences for Intelligent Plant Production System, including:

• Application of probe electrospray ionization for biological sample measurements
• Crop status monitoring in greenhouse based on measurement of volatile organic compounds emitted by the plants
• Environment control for production of valuable materials using plants in closed plant production systems
• Intelligent bio-production system towards environment protection

It is a closed system controlled environment greenhouse. But unfortunately we can’t take any picture inside. It was used for cultivating tomato plants and cucumber. The humidity was automatically controlled by using water sprayer trhough a small pipe-line at the top part of the greenhouse. The humidity sensor detected the low level of humidity, and it will send information to the sprayer to spray the water accurately to keep the humidity level in balance. The temperature is controlled by using heater and cooler. The harvest process was conducted automatically by using harvesting-robot with modified manipulator and machine vision. Machine vision was used to distinguish between the object (cucumber or tomato) and non-object (leaves), because as we know it is difficult to distinguish green cucumber and green leaves only by using visible light camera. Blue and red LED was used to detect photosynthesis rate, as we know that plants during the photosynthesis process absorb more red wavelength rather than other wavelength. My idea is introducing Neural Network based irrigation system….so that the automatic control for water requirements can be much better.

About Ehime Prefecture:

Surrounded by the calm Seto Inland Sea in the north, the Uwa Sea in the west and Shikoku Mountains in the south, Ehime Prefecture, with an area of about 5,676km sq. and a population of about 1,500,000, lies in the northern part of the island of Shikoku, the smallest of the four major islands of Japan.

Its capital, Matsuyama, is the largest city on Shikoku, with a population of over 500,000 and an abundance of historical and cultural resources such as the famous Matsuyama Castle and Dogo Hot Spring Spa, the oldest bathing area in Japan.

Ehime, which sounds feminine to Japanese ears and literally means “beautiful maiden”, is blessed by a mild climate with an average low of 5.3degrees C (42degrees F)in January and an average high of 27degrees C(81degrees F)in August. It offers natural beauty such as the Seto Inland Sea National Park, Ashizuri Uwakai National Park and Mt. Ishizuchi, the highest mountain in western Japan, and attracts millions of tourists annually. It has an average annual rainfall is about 1,300mm, however, most of it falls during the rainy season from mid-June to mid-July. Though there are various explanations concerning its origin, the name Ehime is very old, dating back to ancient myths and legends. Iyo, the old name for Ehime, is said to have first appeared as the name for the island of Shikoku in the “Kojiki” (Record of Ancient Matters) and the “Nihon Shoki” (Chronicles of Japan), both written in the early eighth century. It is also said that the name lyo is derived from lyu, which means hot water, since this area was famous for Dogo onsen (hot spring spa) as far back as the 6th and 7th centuries.

Dogo Onsen

Following the Taika Reform of 645, this area was officially named Iyo-no-kuni. Since the beginning of Ehime’s history, the Seto Inland Sea has played an important role in the life of the people. During the middle ages (after the Taika Reforms),groups of marine guards called suigun appeared. Using their extensive knowledge of tides and currents, they gradually came to dominate over the vital seaborne supply route in the Inland Sea between the westen provinces and Kyoto, then the capital of Japan. Several Suigun active in the region joined together to form a coastal guard unit called the Iyo-Suigun. Their distinguished services in suppressing pirates in the Inland Sea, the Gempei War (late 12th century) and the battles against the Mongol invasion (1274 and 1281), won them a politically important role in the region. They prospered until this area was subjugated by Toyotomi Hideyoshi’s forces in 1585.

During the feudal period, Iyo-no-kuni was cut up into eight fiefs or han known as Iyo- happan by the Tokugawa Shogunate. These were Matsuyama-han, the largest, Ozu-han, Niiya-han, Uwajima-han, Yoshida- han, Imabari-han, Saijo-han and Komatsu- han. Kato Yoshiakira who had fought bravely for Tokugawa Ieyasu in the Battle of Sekigahara (1600), was named lord of Matsuyama-han and established Matsuyama Castle on Katsuyama Hill. While constructing the castle, he arranged the town around the base of the hill, which is Matsuyama’s modern day downtown.

In 1867 the feudal age ended, followed by a restoration of imperial rule. In 1868 the new era was named Meiji and the new government began centralizing its administrative power by abolishing the han and feudal class system in 1871. The Iyo-happan became eight prefectures. Finally in 1873, these prefectures were combined to make Ehime Prefecture, and Prefectural offices were set up in Matsuyama.

My Presentation:

Micro Precision Irrigation System for Sunagoke Moss Production

1. Introduction

This paper describes the design and development of precision irrigation control method by incorporating color, morphology and RGB color co-occurrence matrix (CCM) textural features. Feature selection techniques have become an apparent need in many bioinformatics applications. The objective of this study was to develop neural network based irrigation control model for Sunagoke moss using color, morphology and textural features with the application of feature selection techniques. This system is helpful to explore the new way of water spraying in moss plant factories based on computer vision.

2. Material and Methods

A visible light camera was used to acquire Sunagoke moss images. In this study, samples of cultured Sunagoke moss were used. Growth chamber was used to control optimum growth environment parameters (Hendrawan and Murase, 2008). The temperature (T) in growth chamber was regulated at 15^oC. Humidity (H) was regulated at 65%. Water was given in the amount of 1 gg^-1, 2 gg^-1, 3 gg^-1 and 4 gg^-1. The light was controlled at 30 kflux (12 hours) and the CO₂ gas was controlled at 400 ppm.

Fig. 1 shows that image information were driven to a feature extraction module, where color, morphology and textural features data were obtained. The RGB color components included red index, green index, blue index, red mean value, green mean value, blue mean value and green per red ratio. Morphological features extracted included browning area index, green canopy index and perimeter index. Ten textural features were calculated using RGB CCM, including entropy, energy (angular second moment), contrast, homogeneity, sum mean (mean), variance, correlation, maximum probability, inverse difference moment and cluster tendency. Then, the reduced feature sets obtained after proper feature selection in the feature selection module. The predictor module was generated by applying Back-Propagation Neural Network (BPNN) to different features subsets as the input.

Figure 1. Generic design of noninvasive water content prediction

3. Results and Discussion

Figure 2 shows that water content influences photosynthesis (fig. 2a), color features (fig. 2b), morphological features (fig. 2c) and textural features (fig. 2d). The entire image features and photosynthesis reached the peak point at the water content of 2 gg^-1– 2.5 gg^-1. The result has shown that the combination of image features (color, morphology and textural features) can be used as noninvasive method to monitor the change of water content in Sunagoke moss.

Figure 2. (a) Photosynthesis, (b) color, (c) morphology and (d) CCM textural features

Neural network model performance was tested successfully to describe the relationship between water content and image features. The architecture of BPNN consisted of some combinations of image features as inputs which were selected by features selection techniques, one hidden layer and one output layer respectively. The 124 set data were used in training and inspection data. The training converged after approximately 20.000 iterations, learning coefficient of 0.1 and momentum of 0.9. The mean value of training and inspection root mean square error value were 1.2 x 10^-2 and 0.8 x 10^-2 respectively. For the purpose of feature selection, filter method was used and it showed that feature subset selection had a positive effect on the performance of BPNN on predicting water content.

4. Conclusion

A noninvasive measurement method for Sunagoke moss water content has been suggested. It was concluded that: 1) color, morphology and textural features can be used to monitor the change of Sunagoke moss water content; 2) BPNN method has successfully described the relationship between water content and image features; 3) feature subset selection enhanced the performance of BPNN by improving accuracy; 4) through the application of this system, the water consumption during irrigation process can be minimized and the Sunagoke moss product can be uniform.

5. References

Hendrawan. Y and H. Murase. 2008. Intelligent irrigation control using color, morphological and textural features in Sunagoke Moss. An ASABE Annual International Meeting, Providence, USA, June 29-July 2, paper number: 083858.

Hendrawan. Y and H. Murase. 2008. Water irrigation control for Sunagoke Moss using intelligent image analysis. 17^th IFAC World Congress, Seoul, Korea, July 6-11, paper number: 2723.

Filter techniques assess the relevance of features by looking only at the intrinsic properties of the data. In most cases a feature relevance score is calculated, and low scoring features are removed.

Afterwards, this subset of features is presented as input to the prediction/classification algorithm. Advantages of filter techniques are that they easily scale to very high-dimensional datasets, they are computationally simple and fast, and they are independent of the prediction/classification algorithm. As a result, feature selection needs to be performed only once, and then different classifiers/predictors can be evaluated. A common disadvantage of filter methods is that they ignore the interaction with the classifier/predictor and that most proposed techniques are univariate. This means that each feature is considered separately, thereby ignoring feature dependencies, which may lead to worse classification/prediction performance when compared to other types of Feature Selecition (FS) techniques. In order to overcome the problem of ignoring feature dependencies, a number of multivariate filter techniques were introduced, aiming at the incorporation of feature dependencies to some degree.

The Chi-Squared (X²) method evaluates features individually by measuring their chi-squared statistic with respect to the classes (Liu et al., 2002). For numeric attribute, the method first requires its range to be discretized into several intervals using, for example, the entropy-based discretization method. After calculating the X² value of all considered features, we can sort these values with the largest one at the first position, as the larger the X² value, the more important the feature is. The X² value of an attribute is defined as:

where: m is the number of intervals, k the number of classes, A_ij the number of samples in the ith interval, jth class, Ri the number of samples in the ith interval, C_j the number of samples in the jth class, N the total number of samples, and E_ij the expected frequency of A_ij.

MI function is suitable for assessing the information content of features (Gomez-Sanchis et al., 2008). We evaluated the mutual information (I) function for each feature and class variable, and the features with the highest mutual information function value were selected. The mutual information of two random variables x and y can be viewed as a quantity measuring the mutual dependence of the two variables (Shannon, 1948). The mutual information is widely used in applications area of FS for machine learning.

A probabilistic model of a nominal valued feature Y can be formed by estimating the individual probabilities of the values y ∈ Y from the training data. If this model is used to estimate the value of Y for a novel sample, then the entropy of the model is the number of bits it would take, on average, to correct the output of the model. Entropy is a measure of the uncertainty or unpredictability in a system. The entropy of Y is given by:

If the observed values of Y in the training data are partitioned according to the values of a second feature X, and the entropy of Y with respect to the partitions induced by X is less than the entropy of Y prior t partitioning, then there is a relationship between features Y and X. Equation below gives the entropy of Y after observing X.

where: p(y|x) is the joint probability distribution function of x and y, and p(x) is the marginal probability distribution functions of x respectively.

The amount by which the entropy of Y decrease reflects additional information about Y provided by X and is called the information gain, or alternatively, mutual information (MI) which is given by:

gain = H(Y) – H(Y|X)

= H(X)-H(X|Y)

= H(Y) + H(X) – H(X,Y)

Information gain is a symmetrical measure that is, the amount of information gained about Y after observing X is equal to the amount of information gained about X after observing Y. Symmetry is a desirable property for a measure of feature-feature inter-correlation to have. Unfortunately, information gain is biased in favour of features with more values. Furthermore, the correlations should be normalized to ensure they are comparable and have the same affect. Symmetrical uncertainty compensates for information gain’s bias towards attributes with more values and normalizes its value to the range (0,1):

The Correlation-based Feature Selection (CFS) method is another approach to FS. Rather than scoring (and ranking) individual features, the method scores (and ranks) the worth of subsets of features. As the feature subset space is usually huge, CFS uses a best-first-search heuristic. This heuristic algorithm takes into account the usefulness of individual features for predicting the class along with the level of inter-correlation among them. CFS first calculates a matrix of feature-class and feature-feature correlations from the training data. If the correlation between each of the components in a test and the outside variable is known, and the inter-correlation between each pair of components is given, then the relation between a composite test consisting of the summed components and the outside variable can be predicted from:

where: Merit_S is the heuristic merit of a feature subset S containing k features. r_cf is the average feature-class correlation, and r_ff is the average feature-feature inter-correlation.

From the equation above, it shows that the correlation between a composite and an outside variable is a function of the number of component variables in the composite and the magnitude of the inter-correlations among them, together with the magnitude of the correlations between the components and the outside variable. Entering two illustrative values for r_cf and allowing the values of k and r_ff to vary, the formula is solved for Merit_S. The conclusions of this equation can be drawn:

• The higher the correlations between the components and the outside variable, the higher the correlation between composite and the outside variable
• The lower the inter-correlations among the components, the higher the correlation between the composite and the outside variable.
• As the number of components in the composite increases, the correlation between the composite and the outside variable increases.

Another method is using linear regression (LR). LR analyzes the relationship between two variables, X and Y. In general, the goal of LR is to find the line that best predicts Y from X. LR does this by finding the line that minimizes the sum of the squares of the vertical distances of the points from the line. The LR is defined as:

where: the coefficients a and b are determined by the condition that the sum of the square residuals is as small as possible.

The results of X², MI, CFS and LR were feature ranking which ranked the features based on the relevancy to the predicted variable (water content). Back-propagation Neural Network (BPNN) was used to train selected features based on the ranking number to find relevant combination of features for predicting water content.

The results of FS are described as following. First of all, in most cases, the accuracy of prediction performance using five proposed FS method were greatly improved. It was shown that there were some improvements in case of number of features used and the prediction error between methods using FS and method without FS.

Among the filter methods, X² method got the highest performance for prediction with the validation error 0.009 and 31 features selected followed by MI with testing error 0.018 and 35 features selected.

]]> http://samuraielit.wordpress.com/2008/11/25/jsabees-rihga-royal-hotel-sakai-japan-2007/ Tue, 25 Nov 2008 09:36:00 +0000 samuraielit http://samuraielit.wordpress.com/2008/11/25/jsabees-rihga-royal-hotel-sakai-japan-2007/

……It was a conference held in Rihga Royal Hotel, Sakai, Osaka, Japan in June 2007. I was interested in some fields of study such as: machine vision for precision irrigation, phytotron for closed greenhouse system, plant factory for optimizing environmental growth of plants, etc. I’m introducing fuzzy subtractive clustering in my presentation…

Color Measurement for Estimating Growth Environment of Moss Using Artificial Neural Network

Introduction

The number of application using machine vision and digital image processing techniques in the agricultural sector are increasing rapidly. The quantification of the visual properties can play an important role to improve and automate agricultural management tasks. Non-destructive methods to analyze the growth and development of plants are becoming common and being applied in practice with the development of computers and electronics devices. Simple visible light digital cameras offer a potential for expanded forms of plant ecological research. Using a visible light digital camera can be used as a model for simple field estimations of photosynthesis and carbon gain.

Sunagoke moss Rhacomitrium canescens has been utilized as an active greening material to mitigate the urban heat island effect. The method of image processing of moss was developed, in order to obtain the information about the color appearance based on environment status to be used in intelligent image processing system to determine optimum photosynthesis.

Objectives

The general aim of the project is to investigate the possibility of using intelligent image processing as a tool to facilitate optimum growth environment for photosynthesis management of moss. This has been subdivided into three objectives i.e. developing image processing techniques to determine optimum growth environment of moss based on color appearance, clustering color appearances using Fuzzy Subtractive Clustering (FSC) and modeling the relationship between growth environment parameters and color appearances using an Artificial Neural Network (ANN).

Methods

Moss images were taken through the digital camera and processed by image processing algorithm with image resolution of 500×344 pixels. In this study, samples of cultured Sunagoke moss Rhacomitrium canescens were used. Growth chamber was used to control different variations of environment parameters. The temperature (T) in growth chamber was regulated at 10, 15, or 20 ^oC. Humidity (H) was regulated at 50, 65, or 80%. Water was given in the amount of 1 gr/gr or 2 gr/gr. The light was controlled at 30 kflux (hours of light=12h) and the CO₂ gas was controlled at 400 ppm. The RGB factors i.e. red index, green index, blue index, red mean value, green mean value, and blue mean value were analyzed using Image Processing software built in Visual Basic.

Green area index is defined as total percentage of RGB pixel value which the green color level is up to the threshold points (R=149, G=138, and B=89). These points have been tested effectively to differentiate green area pixel from background and browning area pixel. Browning area index is defined as total percentage of RGB pixel value which is exist in the threshold points (R=70-130 and G=70-100), and it has been tested effectively to measure browning area of cultured Sunagoke moss.

The humidity level was found to be correlated with average green index (r²=0.88), green mean value (r²=0.93), green area index (r²=0.93) and browning area index (r²=0.94).

Result

From the observation, averagely the highest point of the average green index (48.9%) and green mean value (168.75) was found at (T=15-20^oC, H=65%, and water=2gr/gr). the highest increase of green area index (6.87%) was found at (T=20 ^oC , H=65%, and water=2gr/gr). The higher amount of humidity and water content makes the higher value of average green index, green mean value, and green area index. The higher amount of humidity and water content makes the browning area index getting decrease.

In this study, we selected FSC approach to determine cluster centers from the inputs i.e. average RGB index and RGB mean value using parameter radius=0.4. This value is the maximum distance between any two points within the same cluster, yet less than the distance between any two points from different clusters where each point belongs. The multiplier squash factor value=1.25. The criteria for cluster center consideration are based on acceptance and rejection ratios. Acceptance ratio can be determined by fractions of the potential first cluster center, above which another data point will be accepted. Rejection ratio is the condition to reject a data point to be a cluster center, which is obtained from fractions of the potential first cluster center, below which a data point will be rejected as a cluster center. We chose 0.5 as the acceptance ratio for the first cluster center. We chose the rejection ratio 0.15 to derive other cluster centers. We found 3 clusters of color features. From these 3 clusters we can analyze that generally the first cluster has lower value of average green index and green mean value (low humidity and low water content), the second cluster has medium value of average green index and green mean value, and the third cluster has higher value of average green index and green mean value (high humidity and high water content).

Backpropagation neural networks were used to develop the relationship between growth environment parameters with color clusters, green area index and browning area index. These neural networks used temperature, humidity, water, and time as their inputs. Using color clusters as output parameters, a minimum learning error of 5.46×10^-2 was attained at convergence of the ANN training and validation error of ±1.94×10^-2 was attained at ANN validation. Using green area index as output parameters, a minimum learning error of 5.32×10^-2 was attained at convergence of the ANN training and validation error of ±8.66×10^-2 was attained at ANN validation. Using browning area index as output parameters, a minimum learning error of 7.65×10^-2 was attained at convergence of the ANN training and validation error of ±9.05×10^-2 was attained at ANN validation. ANN model performance was tested successfully to describe the relationship between growth environment parameters and color appearance using backpropagation supervised learning method and inspection data.

Conclusion

In this study, we found that visible light camera can visualize growth environment condition using visible color appearance. ANN has successfully described the relationship between growth environment and color appearance. We envision in our future work concerning the combination of visible light image processing and infrared camera. Using the combination of visible light camera features and infrared features can make better observation to visualize the relationship between growth environment condition and optimum photosynthesis process of moss.

The training of a neural network facilitates it to learn to reproduce a mapping from a set of input vectors and the corresponding set of desired output vectors.

The input and output vectors are presented via the input and output layers, respectively. The input layer fans out the input data without making calculations. The data flow along the connection towards the hidden and the output layers. Each unit in the hidden layer transforms the incoming data by executing specified functions. It then outputs the transformed data to the next layer. Each unit in the output layer makes a similar transformation on the data from the hidden layer and, optionally, from the input layer. The final result is that the input vector is mapped into some corresponding output vector at the output layer. If learning is facilitated, the actual output vector is compared with the desired output vector, and the error between the two vectors is calculated. The error values are then used to calculate the weights for all output and hidden layer processing units and thereby reduce the error in the network output. This process is repeated until the mapping has been trained to the desired accuracy or until it appears that the network has learned as well as it can. For a given set of training data, a particular set of weights will result in some degree of mapping accuracy. The purpose is to find a set of weights that will minimize error. Input features were for the models were image features including color, morphology and textural features, respectively. A three layer structure shown in Figure 1 was adapted for all the models.

Fig. 1 Architecture of neural network (BPNN) models.

During the forward propagation, the network transfers its input data as shown by equation (1), where x_i is the activation level of unit i, and v_ij is the synapse weight from unit i to unit j (unit i and j are in the input and hidden layers, respectively), and z_in_j is the weighted sum of inputs to the j^th unit in the hidden layer.

equation (1)

The hidden unit j has a sigmoid transfer function that computes the output z_j shown by equation (2). The output z_j becomes the next input for the output unit k, which has a summation through another synapse weight w_jk, i=1,2,…m, j=1,2,…h and k=1,2,…n. The weighted sum of inputs y_in_k in equation (3) received by the output unit k undergoes a transfer function y_k, expressed by equation (4).

equation (2)

equation (3)

equation (4)

The backward propagation involves calculating, for all processing units, error values that are used during synapse weight adjustments. The calculations being at the output layer and progress backward through the network to the input layer. Equation (5), respectively, compute the error values δ_k for unit k in the output layer and unit j in the hidden layer. In the equations, t_k and y_k are the target and the output values of unit k, respectively, f’(x) is the derivative of the sigmoid function f, and y_in_k is the weighted sum of inputs to j and k, respectively. For the hidden layer, a weighted sum is taken of the values of δ for all units that receive output from unit j.

equation (5)

Each synapse weight is adjusted by taking into account the δ value of the unit that receives input from that interconnection. The weight adjustment is based on equation (6 and 7), where α is the learning rate which is the error value of the target unit, and z_j is the output value of the originating unit.

equation (6)

equation (7)

In every hidden unit z_j ­, j= 1,2,…h we can calculate error information in equation (9) by using weighted input in equation (8) and its derivative value.

equation (8)

equation (9)

The corrected weight in hidden layer can be calculated using equation (10) and equation (11) which will be used to improve the synapse weight v_ij.

equation (10)

equation (11)

Every output unit y_k, k=1,2,…n will improve bias value and synapse weight w_jk value j=1,2,…h as shown in equation (12).

equation (12)

Every hidden unit z_j, j=1,2,…h will improve bias value and synapse weight v_ij value i=1,2,…m as shown in equation (13).

equation (13)

The number of neurons in the input layer was determined by the number of input features. The number of neurons in the hidden layer was determined by trial and error. The output layer consisted of a single neuron. The MSE given by equation 14 was used as the performance criterion for the models:

equation (14)

where: N_n is number of input feature vectors, S_i is the water content predicted by the BPNN model, and St_i is the target of water content.

In this study, the BPNN theory, generally accepted as a useful tool for the recognition of various patterns, and image processing technique were used for the development of the desired software. A three layer BPNN model with default number of neurons in one hidden layer was used in this study. The sample’s color, morphology and textural features were used to develop BPNN models for water content estimation. The 212 samples data of Sunagoke moss were randomized and divided into two parts which were 75% data as training set and 25% data as validation set. The models were trained in 400 iterations.

Neural network model performance was tested successfully to describe the relationship between water content and image features. Comparison analysis as shown in table 1 was conducted to find the best model to determine water content using color, morphology and textural features and three layers BPNN. The validation-set results show that BPNN method can estimate water content of Sunagoke moss better than SPC method using 4 image features i.e: (1) green/red ratio mean value; (2) blue index; (3) blue mean value and (4) green index with the reliability improvement value of 82.33%. The performance of BPNN using all features (106 features) has shown good result in training but not so well in validation. The combination of Saturation_HSV CCM textural features performed better in detecting water content than other models respectively based on validation-set with the MSE of 0.0226. The most five contributing parameters for the water content estimation using BPNN based on validation-set were: (1) Saturation_HSV CCM textural features (MSE: 0.0226); (2) Saturation_HSL CCM textural features (MSE: 0.0239); (3) Red CCM textural features (MSE: 0.0261); (4) Blue CCM textural features (MSE: 0.0261) and (5) Green CCM textural features (MSE: 0.0273).

Table 1 Comparison analysis using SPC and BPNN.

Selecting features that are suitable for an application is one of the most important parts in solving the problem.

Feature subset selection is the process of identifying and removing as much irrelevant and redundant information as possible. This reduces the dimensionality of the data and may allow learning algorithms to operate faster and more effectively. In some cases, accuracy on future classification can be improved; in others, the result is a more compact, easily interpreted representation of the target concept.

FS algorithms perform a search through the space of feature subsets, and, as a consequence, must address four basic issues affecting the nature of the search (Langley, 1994):

1. Starting point. Selecting a point in the feature subset space from which to begin the search can affect the direction of the search. One option is to begin with no features and successively add attributes. In this case, the search is said to proceed forward through the search space. Conversely, the search can begin with all features and successively remove them. In this case, the search proceeds backwards through the search space. Another alternative is to begin somewhere in the middle and move outwards from this point.

2. Search organization. An exhaustive search of the feature subspace is prohibitive for all but a small initial number of features. With N initial features there exist 2^N-1 possible subsets. Heuristic search strategies are more feasible than exhaustive ones and can give good results, although they do not guarantee finding the optimal subset.

3. Evaluation strategy. How feature subsets are evaluated is the single biggest differentiating factor among feature selection algorithms for machine learning.

4. Stopping criterion. A feature selector must decide when to stop searching through the space of feature subsets. Depending on the evaluation strategy, a feature selector might stop adding or removing features when none of the alternatives improves upon the merit of a current feature subset. Alternatively, the algorithm might continue to revise the feature subset as long as the merit does not degrade. A further option could be to continue generating feature subsets until reaching the opposite end of the search space and then select the best.

As many pattern recognition techniques were originally not designed to cope with large amounts of irrelevant features, combining them with FS techniques has become a necessity in many applications. The objectives of FS are manifold, the most important ones being: a) to avoid over-fitting and improve model performance, i.e. prediction performance in the case of supervised classification and better cluster detection in the case of clustering, b) to provide faster and more cost-effective models, and c) to gain a deeper insight into the underlying processes that generated the data.

FS techniques can be organized into three categories depending on how they combine the FS search with the construction of the classification/prediction model: filter methods, wrapper methods and embedded methods. In this study we focus to solve FS problem using filter and wrapper methods. Filter techniques assess the relevance of features by looking only at the intrinsic properties of the data. Whereas filter techniques treat the problem of finding a good feature subset independently of the model selection step, wrapper methods embed the model hypothesis search within the feature subset search. In a third class of FS, termed embedded techniques, the search for an optimal subset of features is built into the classifier construction, and can be seen as a search in the combined space of feature subsets and hypotheses. Just like wrapper approach, embedded approaches are thus specific to a given learning algorithm. Embedded methods have the advantage that they include the interaction with the classification/prediction model, while at the same thing being far less computationally intensive than wrapper methods.

Pattern recognition, one of the fields of machine learning, is defined as the imposition of identity on input data, such as speech, images, or a stream of text, by the recognition and delineation of patterns it contains and their relationship.

Methods of pattern recognition are used to classify recurring patterns in data on the basis of either a priory knowledge or information directly extracted from the data.Thus, the resulting learning strategy is characterized as either supervised or unsupervised learning. Normally, pattern recognition involves gathering observations on certain objects in a standardized manner and classifying them on the basis of a formerly agreed set of object features.

In the past decade, various researchers have used image processing and pattern recognition techniques in agricultural applications. The underlying approach for all of the pattern recognition techniques is the same. First, images are acquired from the environment using an analog, digital, or video camera. Then, image processing and image analysis techniques are applied to extract useful features that are necessary for further analysis of these images. Afterwards, discriminant techniques are employed to classify the images, using such techniques as parametric or non-parametric statistical and neural networks (NN), which are available to address the specific problem at hand. The selection of the image processing techniques and the classification strategies are important for the successful implementation of any machine vision system, and are typically application dependent. In the past few years, there have been numerous studies of the application of NN in the research related to agriculture. NN is an aspect of artificial intelligence (AI) research.

So she said what’s the problem baby
What’s the problem I don’t know
Well maybe I’m in love (love)
Think about it every time
I think about it
Can’t stop thinking ’bout it
- Accidentally In Love

KENAPA NERAKA!!!! Beeuuhh, barusan tau dari kholes . Ternyata senen besok ada kelas REGIS : Research Group On Intelligent System jam setengah 4 sore. Padahal harusnya seperti yang sudah distate oleh advisor REGIS (Pak Ruly Soelaiman Skom Mkom), pertemuan selanjutnya setiap korban pengikut REGIS ini diharapkan sudah membawa (dan membaca tentunya) paper untuk didiskusikan belum. Langsung deh blingsatan dan gelagapan. Mulai obok obok IEEE dan google deh saya.

Mulailah dengan sesuatu yang anda minati, begitu nasehat dari advisor REGIS. Karena bingung apa yang sebenarnya saya minati , makanya saya gak mulai mulai (BEUUH!!). Setelah memaksa dan menarik diri saya sendiri, mulai deh saya browsing. Mulai untuk mencari cari topik dalam belantara Intelligent System yang penuh dengan rumus rumus liar, pasir hisap integral differensial dan angka angka haus darah. Akhirnya ada satu topik yang menarik buat saya, web 3.0.

Generasi Web, diambil dari sini

Merangkum apa yang sudah tertulis di wikipedia, web 3.0 sebenarnya hanya terminologi yang digunakan oleh orang orang yang berkecimpung di dunia web untuk menggambarkan generasi baru dari dunia WWW. Beberapa inovasi yang diprediksikan dapat menjadi pemicu alih generasi web dari 2.0 ke 3.0 adalah :

• Web Semantik (saya masih belum ngeh masalah ini)
• Mashups (menggabungkan konten dari macem macem web untuk personalisasi)
• Service Oriented Architecture (WSDL, SOAP, Web Service, BPEL!! Ampun DJ!! calon TA gua!!!)
• 3D Web Apps (kayak SecondLife gtu mungkin)
• Data Web & Web Mining (Web sebagai database)
• Intelligent Web (Nah ini yang saya pengen bahas, blum ngerti tapi)

Untuk paper REGIS, masih nyari nyari yang masalah Intelligent Web ini. Setelah beberapa puluh menit dihabiskan. Cuman nemu 2 abstrak paper doank. Karena dua paper ini ada di jurnal berbayar, jadi bingung gimana cara nya donlot.

Dua paper yang saya temukan ini antara lain :

• Adapting agent communication languages for semantic Web service inter-communication
• A fuzzy web intelligence application: web-based fuzzy expert agents for online learning evaluation
  

Yang pertama mbahas masalah komunikasi data untuk web semantik.Yang kedua implementasi logika fuzzy untuk penentuan keputusan berdasarkan penalaran halus (soft reasoning) berbasis web. Tenang bro, kedua duanya saya juga sama sekali gak NGERTI BLAS. Yang kedua kayaknya yang lebih masuk ke Intelligent System. Kata kata “fuzzy” udah bikin bergidik, untungnya di abstrak paper tersebut disebutkan bahwa implementasinya menggunakan servlet dan JDBC. Lumayanlah buat memotivasi sedikit.

Sementara ini masih menceburkan diri ke wikipedia untuk mencoba mengerti beberapa istilah istilah bahasa klingon yang menghambat proses untuk memahami kedua paper laknat tersebut. Beberapa istilah tersebut antara lain:

• Fuzzy Logic
• Soft Computing
• Orchestration
• Context Aware
• Semantic Web
• Ontology
• Web Mining

Tampaknya malem minggu ini saya habiskan dengan brain boosting lagi deh. Kapan gak jomblonya klo kayak gini terus T_T. Semangat temen temen!!!

just open up your heart and let the sun shine in
- frente

Tulisan terkait

• REGIS Session: Research What, How and Why
• Intelligent System Research Group
• 3 Bidang Minat
• Malem minggu: Boosting my brains

no, what do you own the world?
How do you own disorder, disorder,
Now, somewhere between the sacred silence,
Sacred silence and sleep,
Somewhere, between the sacred silence and sleep,
Disorder, disorder, disorder.
- Toxicity, System Of A Down

Hari kamis kemaren adalah pertemuan pertama dari mata kuliah Intelligent System. Mata kuliah yang saya anggap angker. Dan kebetulan di ampu (di ajar) oleh dosen yang “angker” juga. Sebelumnya saya sebenarnya pengen menghindar dari dosen teresebut, tetapi karena dua kelas beliau semua yang mengajar ya apa boleh buat. Kenapa mata kuliah ini angker?because its all about mathematics and stuff. Bayangin aja gimana kedernya waktu mendengarkan penjelasan beliau mengenai integral tingkat 10. Im really suck at math dan mata kuliah ini bakal jadi mimpi buruk bagi saya.

Setelah beberapa menit pembukaan dan overview mengenai machine learning, neural network dan review paper beliau, ada sesuatu yang benar benar baru saya dengar sepanjuang sejarah perkuliahan saya. Pak Rully Soelaiman, dosen saya ini menawarkan untuk membagi kelas Intelligent System menjadi dua. Kelas reguler dan kelas riset. Klo yang kelas reguler kuliahnya biasa aja, klo yang kelas riset kuliahnya modelny mirip dengan kuliah s2. Setiap mahasiwa diberikan sebuah paper, dan sepanjang kuliahnya itu isinya mbahas paper, presentasi, bikin proposal riset dan segala macem. Dan yang jelas materi papernya gak jauh jauh dari Intelligent System ini. Setelah itu dibagikan kertas pendaftaran, yang tertarik tanda tangan di kertas itu.

Saya tanda tangan di kertas itu. Nekat? Yap!! Ini kontrak mati saya dengan Pak Rully. Seperti yang pernah saya ceritakan pada postingan sebelumnya, seumur hidup saya menghindar dengan sesuatu yang berbau matematis. Ada gejala kejang kejang mirip epilepsi setiap kali saya melihat rumus dan macem macemnya. ketika dihadapkan dengan pilihan seperti ini? apakah saya akan terus lari? Nggak, saya memilih buat stand up and take the stake. Sampai kapan saya harus lari? dan ini kesempatan buat saya untuk melihat sejauh mana sih batas kemampuan saya. Dan saya memprediksikan bakalan berdarah darah di kelas ini, im going to walk thru hell. Saya juga pengen liat, sejauh mana saya bisa bertahan dari hentakan gelombang otak melihat rumus rumus neural network yang bejibun.

Kelas riset masih mulai sekitaran 2 minggu lagi. Masih ada waktu buat mempersiapkan mental, tapi sudah terlambat untuk mundur. Sementara ini masih asyik baca baca masalah SOA, thanx to mas Joni Farizal. Ada yang sempet nanya, “Gak jadi E business mbel?” wekekekek kayaknya masih pur, ini kan riset. Dan buat informasi aja, kelas riset cuman sekitar 20 an orang. Hmm kayaknya kita butuh nama deh buat ini. Ada yang punya usul?

Segitu deh laporannya. Tetep semangat!!!

: )