[ieee 2012 first international conference on agro-geoinformatics - shanghai, china...

6
Ant Colony Optimisation based Land Use Suitability Classification Jia Yu* Department of Geography Shanghai Normal University Shanghai, China *Corresponding author, e-mail: [email protected] Yun Chen Land and Water CSIRO Canberra, Australia Jianping Wu, Chang Huang Key Laboratory of Geographic Information Science, Ministry of Education East China Normal University Shanghai, China Abstract—This paper presents a new land use suitability classification (LSC) method on the basis of Ant Colony Optimisation (ACO), which is one kind of AI techniques. ACO algorithm can be used to discover suitability classification rules according to training cases. Classification rules and training cases are all organised in the form of IF-THEN, which generally incorporates practical human knowledge. To implement ACO based LSC, a tool was developed using ArcGIS Engine component in .NET framework. The tool provides some useful functions and interfaces for the integration of spatial data input, sampling of training cases, rule classification discovery and LSC mapping. A case study in the Macintyre Brook Catchment of southern Queensland in Australia is proposed. The tool was used to process land use suitability classification in the study area for irrigated agriculture. The resultant map was then compared with present irrigated land to show spatial distribution of irrigated land suitability and to reveal future potential of land use development in this area. Further analysis was conducted to demonstrate the feasibility of ACO method. The parameter values were adjusted to explore the robustness of parameter settings. We also compared the ACO method with C4.5 which is a kind of decision tree algorithm. It has been found that ACO method can produce simpler rule list with slightly reduced classification accuracy. Therefore, in our point of view, although with it limitation, the ACO method is a practicable and efficient approach, and worth more research. Keywords- ant colony optimisation (ACO); classification rule; land use suitability; GIS I. INTRODUCTION The classification process of land use suitability is an important task in land use evaluation. The most important task of land use suitability classification (LSC) is to classify specific areas of land according to their suitability for defined uses [1]. There are kinds of approaches which have been proposed, such as weighted averaging [2], analytical hierarchy process (AHP) [3-5] and ordered weighted averaging (OWA) [6, 7]. In recent years, artificial intelligence (AI) techniques have been extensively used in LSC, such as artificial neural networks (ANN) [8, 9], decision tree [10], expert systems [11], genetic algorithm (GA) [12] and cellular automata (CA) [13]. This paper adapts an advanced AI technique - ant colony optimisation, into LSC study. It aims to conduct more reasonable spatial LSC results according to better classification rules. For LSC, rule-based classification is a key process to obtain classification rules. It is regularly used to store, distribute, reason about and apply practical problem-solving knowledge for working out complex problems in proper ways [14]. The rule-based classification always translates human knowledge into a conditional statement with the form of “IF…THEN…” One of the advantages of using spatial rule- based classification is that valuable information and knowledge hidden in massive spatial data can be classified. Therefore, this technique has been applied in land use suitability evaluation and land-use planning [14-16]. Ant colony optimisation (ACO) is one of artificial intelligence methods [17-18]. It has been utilised to resolve kinds of problems in real world, such as the scheduling problem [19] and the vehicle routing problem [20]. The concept for ACO was motivated by the foraging behavior of real ants. Parpinelli et al (2002) proposed the first ACO algorithm for rule-based classification [21]. They put forward an algorithm called Ant-Miner. It was used to extract classification rules from data. After that, some techniques on the basis of Ant-Miner were proposed, for instance, Ant- Miner2 and ACO-Miner [22, 23]. These techniques improved the performance or extend the applicability of Ant-Miner in handling non-spatial problems. But to solve spatial LSC problems, these applications are not able to be utilised without adaption and modification. There exist two major challenges: (1) Import and export of spatial data makes sampling process and rule classification more difficult; (2) Spatial problems always involve massive data, therefore, “simpler classification rules” is also important comparing with “better performance”.

Upload: chang

Post on 08-Dec-2016

215 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

Ant Colony Optimisation based Land Use Suitability Classification

Jia Yu* Department of Geography

Shanghai Normal University Shanghai, China

*Corresponding author, e-mail: [email protected]

Yun Chen Land and Water

CSIRO Canberra, Australia

Jianping Wu, Chang Huang Key Laboratory of Geographic Information Science,

Ministry of Education East China Normal University

Shanghai, China

Abstract—This paper presents a new land use suitability classification (LSC) method on the basis of Ant Colony Optimisation (ACO), which is one kind of AI techniques. ACO algorithm can be used to discover suitability classification rules according to training cases. Classification rules and training cases are all organised in the form of IF-THEN, which generally incorporates practical human knowledge. To implement ACO based LSC, a tool was developed using ArcGIS Engine component in .NET framework. The tool provides some useful functions and interfaces for the integration of spatial data input, sampling of training cases, rule classification discovery and LSC mapping. A case study in the Macintyre Brook Catchment of southern Queensland in Australia is proposed. The tool was used to process land use suitability classification in the study area for irrigated agriculture. The resultant map was then compared with present irrigated land to show spatial distribution of irrigated land suitability and to reveal future potential of land use development in this area. Further analysis was conducted to demonstrate the feasibility of ACO method. The parameter values were adjusted to explore the robustness of parameter settings. We also compared the ACO method with C4.5 which is a kind of decision tree algorithm. It has been found that ACO method can produce simpler rule list with slightly reduced classification accuracy. Therefore, in our point of view, although with it limitation, the ACO method is a practicable and efficient approach, and worth more research.

Keywords- ant colony optimisation (ACO); classification rule; land use suitability; GIS

I. INTRODUCTION The classification process of land use suitability is an

important task in land use evaluation. The most important task of land use suitability classification (LSC) is to classify specific areas of land according to their suitability for defined uses [1]. There are kinds of approaches which have been proposed, such as weighted averaging [2], analytical hierarchy process (AHP) [3-5] and ordered weighted averaging (OWA) [6, 7]. In recent years, artificial intelligence (AI) techniques have been extensively used in LSC, such as artificial neural networks

(ANN) [8, 9], decision tree [10], expert systems [11], genetic algorithm (GA) [12] and cellular automata (CA) [13]. This paper adapts an advanced AI technique - ant colony optimisation, into LSC study. It aims to conduct more reasonable spatial LSC results according to better classification rules.

For LSC, rule-based classification is a key process to obtain classification rules. It is regularly used to store, distribute, reason about and apply practical problem-solving knowledge for working out complex problems in proper ways [14]. The rule-based classification always translates human knowledge into a conditional statement with the form of “IF…THEN…” One of the advantages of using spatial rule-based classification is that valuable information and knowledge hidden in massive spatial data can be classified. Therefore, this technique has been applied in land use suitability evaluation and land-use planning [14-16].

Ant colony optimisation (ACO) is one of artificial intelligence methods [17-18]. It has been utilised to resolve kinds of problems in real world, such as the scheduling problem [19] and the vehicle routing problem [20]. The concept for ACO was motivated by the foraging behavior of real ants. Parpinelli et al (2002) proposed the first ACO algorithm for rule-based classification [21]. They put forward an algorithm called Ant-Miner. It was used to extract classification rules from data. After that, some techniques on the basis of Ant-Miner were proposed, for instance, Ant-Miner2 and ACO-Miner [22, 23]. These techniques improved the performance or extend the applicability of Ant-Miner in handling non-spatial problems. But to solve spatial LSC problems, these applications are not able to be utilised without adaption and modification. There exist two major challenges: (1) Import and export of spatial data makes sampling process and rule classification more difficult; (2) Spatial problems always involve massive data, therefore, “simpler classification rules” is also important comparing with “better performance”.

Page 2: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

This study conducts the ACO method to handle the spatial LSC problems. We developed a GIS-based land use suitability assessment tool to aggregate numerous attributes and discover appropriate rules. A case study of the Macintyre Brook catchment in southern Queensland of Australia is presented to demonstrate the method.

II. ACO FOR RULE BASED CLASSIFICATION

A. Rule Structure In general, the structure of classification rules can be

expressed in the form as follows:

IF <Antecedent> THEN <Consequent>

The <Antecedent> of the rule includes a number of conditions, commonly connected by a logical conjunction operator AND. The <Consequent> part presents the classification result of the rule.

Figure 1. Data structure for classification rule discovery

B. Data Structure of Rule Discovery The data structure for discovering rules is motivatied from

the foraging behavior of ant colonies in the real world. To simulate the ant behaviors and build artificial foraging paths, the data structure of rule discovery as Fig. 1 is constructed. All the artificial ants move from the Start node, which is regarded as the virtual nest. Below the nest are a number of attributes, and every attribute has several values. An attribute is symbolised by Attributei, in which i is the sequence number of the attribute. Vij is a value belongs to an attribute, I is the sequence number of this attribute, and j is the sequence number of value in this attribute. Each discrete attribute has a number of discrete values. Attribute1 and Attribute3 are two discrete attributes. Attribute1 has four discrete values: V1,1, V1,2, V1,3 and V1,4. Attribute3 also has four values: V3,1, V3,2, V3,3 and V3,4. Attribute2 is a continuous attribute which has continuous values V2. The values of it are discretised into three parts by two thresholds: T2,1 and T2,2. The three parts of the values are, Attribute2< T2,1 , T2,1≤ Attribute2< T2,2 and Attribute2≥ T2,2. Below the attributes is the Class. A value of the Class is expressed as Ck where k is the sequence number of the value in the Class. An ant will start from the artificial nest and chooses a value for each attribute. After it goes through all the attributes, it will choose a value for the Class and then take the artificial food. In Fig. 1, there is a path taken

by an ant which is represented with nodes connected by red lines: Start→V1,2→(< T2,1)→V3,3→C3→End. It forms an IF-THEN rule, which has been mentioned in section 2.1. This rule can be expressed as:

IF Attribute1 = V1,2 AND Attribute2 < T2,1 AND Attribute3 = V3,3 THEN Class = C3.

C. ACO Algorithm for Rule Discovery Fig. 2 represents the workflow of the ACO algorithm for

rule discovery. The algorithm begins with obtaining a TrainingSet which containing numbers of training cases. The main loop runs to discover one rule per iteration. A sub-loop is executed to find out classification rules by a number of ants who travel on the path in turn. The main loop will be finished when the number of training cases is less than or equal to a user-specified threshold. At the same time, all the discovered rules will be added to the rule list. Please refer to the paper authored by Parpinelli et al [21] to obtain detailed description of the algorithm.

Figure 2. Workflow of ACO algorithm for rule discovery

III. THE FRAMEWORK OF THE TOOL The architecture of the tool was illustrated in Fig. 3. Firstly,

it reads all attribute layers in raster format. The discrete and continuous attribute layers need to be standardised to the same geographic scale, boundary, cell size and spatial reference.

After the attribute layers are all imported, we collect a certain amount of samples in different spatial locations from each attribute layer so as to build training cases for the training set. Expert knowledge and field survey play an important role during the sampling. Expert knowledge can assist to simplifying and abstracting complex information into data in standard format. The building of the training set is crucial because it straightly makes effect to the precision and quality of the discovered rules.

The tool was designed as a raster-based tool in ArcGIS environment with Microsoft C# .NET computer language and

Page 3: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

ESRI ArcGIS Engine development components. Function modules designed to construct tool framework are as follows:

LSA Attribute Layers(Raster Format)

Discrete Layers

Continuous Layers

Training Set

Training set input

Classification Rules

Discover

ClassifySupport

Unclassified Empty Layer

Sampling

LSA Result Map Layer

Generate

Knowledge and field survey

Preprocess & Integrate

ACO Algorithm

Figure 3. Implementation processing architecture

A. Training Cases Collection Module The function of this module is to collect training cases into

the training set. The collecting process gets the antecedent part of cases from spatial sampling, and extracts the consequent part from the knowledge repository. The knowledge repository built for the tool is a set of local knowledge information of suitability classification of specified study area. The knowledge can be gathered through integrating experts’ experiences and field surveys.

B. ACO Modeling Module This module is designed according to the ACO algorithm.

C. Spatial Data Processing Module This module is on the basis of GIS technical. It reads raster

layers, conducts spatial sampling and applies discovered rules to the study area according to the spatial classification. Finally, it outputs the LSC result. The module can also achieve other tasks, such as result map export, map document saving and classification area statistics.

D. Map Layer Display Module The main task of this module is to visualise spatial data

and relatively completes basic map operations such as zoom in, zoom out, pan and full extent. Once the Spatial data processing module generates the resultant map, this module displays special suitability levels for different class and renders them with themaitic colors.

IV. CASE STUDY To demonstrate the practicability of the proposed method,

a case study in the Macintyre Brook Catchment, southern Queensland, Australia, was conducted to evaluate the land use suitability for irrigated agriculture.

The Macintyre Brook Catchment has an area of 4,200 km2 (Fig. 4). The catchment is characterised by extremely diverse soil types and topography [24], making it suitable for a wide diversity of land use (Fig. 4) and rural production. The Macintyre Brook flows from east to west, along which the main irrigation areas are located [25]. Now about 1.5% of the catchment area is allocated to irrigated cropping, perennial horticulture and sown pastures. The rest agriculture land is dominated by dryland cropping (3%), native pasture grazing country (80%) and State Forest Reserves (15%). The major crops consist of fodder (lucerne), maize, sorghum, peas, and orchard such as peach, plum and apricot [26, 27].

Figure 4. Location and landuse (100m resolution) of Macintyre Brook

A. Generation of Criterion Maps This study adapted the suitability classes which were

recommended by the Food and Agricultural Organisation [1]. Four classification levels are consisted: highly suitable (S1), moderately suitable (S2), marginally suitable (S3) and unsuitable (N). Five attributes were selected, including percent slope (S), soil texture (ST), depth to water-table (DTW), electrical conductivity of groundwater (ECw) and hydraulic conductivity of soil (Ks). The attribute values of S and DTW are continuous values; the others are discrete values which need to be pre-processed.

The attribute layers were normalised in raster format with a cell of 100×100m and given the spatial reference of UTM Zone 56S. Based on the predefined thresholds, discrete attribute layers ST, ECw and Ks are classified into four classes. The values of the thresholds were assigned on the basis of literature review and expert opinions according to broad scale analysis of irrigated cropping in this catchment. Each cell in raster layers was given numerical values 4, 3, 2 or 1, respectively stands for different levels: S1, S2, S3 or N.

Page 4: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

B. Build Training Set To build the training set for the case study, we took 500

samples within the area so as to achieve the generation of the training cases (Table I). Each case in the training set is made up of two parts, antecedent part and consequent part. The cell values sampled from different attribute layers identify the antecedent part of the cases. Field surveyed data form the consequent part of the cases. If it was not easy to determine the consequent part only based on field survey, expert knowledge was then required to be involved.

There are two sampling methods designed in the tool. Some key samples which gathered in the irrigation area and the flood plain were positioned by the function of manual sampling. Other samples were gained by random sampling. The values of the cell sampled generate the antecedent part of the cases. Data collected by field survey and expert knowledge determined the consequent part of the training cases.

TABLE I. EXAMPLES OF TRAINING CASES

Training Cases

S ST DTW ECw Ks Class

1 S3 S2 S1 S3 N S3 2 S1 S1 S1 S3 S2 S2 3 N N S1 S3 N N

C. Rule Discovery and Map Generation The tool was implemented after inputting training cases.

Rules were generated and saved through iterative runs. The discovered rules in the tool for this study area had the format of “ )T(V)(T 2

i1i <≤ S

i , STjV , )T(V)(T 2

k1k <≤ DTW

k , ECwlV ,

KsmV , nC ”, where i, j, k, l, m are the sequence numbers of

attributes and n is the sequence number of the Class. The threshold in the parentheses is optional. It means

)T(V)(T 2i

1i <≤ S

i can have the format of 2

iTV <Si

, 2i

1i TVT <≤ S

i

or SiVT1

i ≤ . If no node of an attribute was selected, the value of that attribute would be assigned a value of -1. For instance, rule “≤3,-1,-1, 2,-1, 3” in the ACO-LSA tool was equivalent to the rule:

IF S≤ 3 AND ECw = 2 THEN Class = 3.

Once all the rules were discovered and added to the rule list, they would be applied to the whole study area with the same order as that they were discovered. After that, a LSC resultant map can be generated on the basis of these classification rules.

D. Result Analysis and Discussion Fig. 4 presents the LSC resultant map. The regions

dominated by highly suitable (S1) are largely distributed on the flood plain of the Macintyre Brook. Unsuitable (N) regions are located in the south-east areas of the catchment. These regions have relatively steep slopes, poorer soil texture, and lower hydraulic conductivity than the flood plain.

Figure 5. Land-use suitability map of Macintyre Brook Catchment of southern Queensland, Australia: LSC resultant map

The simulated S1 regions are overlaid with present irrigated land. The comparison result reveal that just less than 20% of the highly suitable areas have been developed for irrigation production. It means this area has potential to expand the irrigation land use in the future. The spatial distribution of the land use suitability has demonstrated that the impact of the DWT and ECw on the LSC results is not significant. The value of Ks value has a notable impact on the results, because more than 90% S1 and S2 areas are allocated in the regions where the value of Ks is from 0.05 to 2m/d. ST also has remarkable contribution to the classification. More than 50% of the S1 and S2 lands have fine to medium soil texture. As to the slope (S), the spatial analysis also shown that over 60% S1 and S2 regions are of flat slopes (percent slope <2%).

Judge from the above analysis, we come to the opinion that, if local financial budget permits and water supply can be guaranteed for irrigation, the catchment has its potential to expand its irrigation areas. The LSC resultant map (Fig. 4) can be utilised to help establishing new irrigation plans in the future. Local decision makers can refer to this map to decide the schedule of development of irrigation with the integration of political, social and economic factors.

E. Comparison of ACO based LSC with C4.5 ACO based LSC method is compared with C4.5 decision

tree algorithm, which is a well-known and broadly used rule-based classification method. A decision tree is a popular classification method that builds an interpretable model to represent a set of rules. C4.5 is an algorithm used to generate a decision tree developed by Quinlan in 1993 [28]. The discovered rules between ACO based LSC and C4.5 are of the same structure. Therefore, the comparison of them is meaningful. In this study, the same training set is used and the performances of these two methods are evaluated respectively.

Another 500 samples are taken to create the test set to verify the results of both methods and take comparison of classification accuracy (Table II). The comparing results illustrated that ACO based LSC and C4.5 have same

Page 5: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

classification accuracy for S1. As to the other classes, the accuracies are slightly lower, especially S3 and N. Referring to the simplicity of discovered rule list, ACO based LSC discovered 18 rules, whereas C4.5 discovered 40 rules (Table II). The length of rule list of ACO based LSC is less than half of that of C4.5. It means ACO based LSC is much simpler and easier to be understood. The simplicity of the rule list for LSC is significant. It can help relative personnel understand the discovered knowledge better and easier to explain the critical attributes in LSC.

TABLE II. CLASSIFICATION ACCURACY AND NUMBER OF RULES DISCOVERED OF ACO BASED LSC AND OF C4.5

Acurracy (%) Rules S1 S2 S3 N Discovered

ACO 99.28 93.18 97.84 97.78 18

C4.5 99.28 95.45 98.56 98.89 40

V. CONCLUSIONS In this study, a new method of Ant Colony Optimisation

(ACO) based land use suitability classification (LSC) is proposed which support continuous attributes as well as discrete ones. In comparison with other land use evaluation methods, this method utilises easy-understood rules for assessment, while most other approaches are on the basis of mathematical equations, which are implicit for users. The IF-THEN expression of rules is close to natural language. The rules and the rule list are simple and easy for users to interpret. Moreover, compared with other common rule-based classification methods, the ACO algorithm has acceptable classification accuracy and much simpler rule list. In conclusion, we believe the ACO based LSC method is a practicable and efficient approach which worth more research.

VI. ACKNOWLEDGMENTS The authors would like to express great appreciations to the

Department of Natural Resources and Water of the Queensland Government for providing data of the study which was supported by the System Harmonisation program of the Cooperative Research Centre for Irrigation Futures (CRCIF), Australia.

REFERENCES [1] Food and Agriculture Organization of the United Nations (FAO). “A

Framework for Land Evaluation”, Soils Bulletin. No.32. Rome: FAO, 1976.

[2] J. Malczewski, “GIS and Multicriteria Decision Analysis”, New York: John Wiley & Sons, Inc, 1999.

[3] T.L. Saaty, “The analytic hierarchy process: planning, setting priorities, resource allocation” (287pp), New York: McGraw Hill International Book Cooperation, 1980.

[4] T. Cengiz and C. Akbulak, “Application of analytical hierarchy process and geographic information systems in land-use suitability evaluation: a case study of Dümrek village (Çanakkale, Turkey)”. International Journal of Sustainable Development & World Ecology, vol. 16 (4), 2009, pp. 286-294.

[5] Y. Chen, J. Yu and K. “Shahbaz, Spatial Sensitivity Analysis of Multi-Criteria Weights in GIS-based Land Suitability Evaluation”, Environmental modelling and Software. vol. 25(12), 2010, pp. 1582-1591.

[6] J. Malczewski, T. Chapman, C. Flegel, D. Walters, D. Shrubsole and M.A. Healy, “GIS multicriteria evaluation with ordered weighted averaging (OWA): case study of developing watershed management strategies”. Environment and Planning A, vol. 35 (10), 2003, pp. 1769-1784.

[7] J. Malczewski, “Ordered weighted averaging with fuzzy quantifiers: GIS-based multicriteria evaluation for land-use suitability analysis”. International Journal of Applied Earth Observations and Geoinformation, vol. 8 (4), 2006, pp. 270-277.

[8] D.Z. Sui, “Integrating neural networks with GIS for spatial decision making”. Operational Geographer, vol. 11 (2), 1993, pp. 13-20.

[9] C.J.A. Bradshaw, L.S. Davis, M. Purvis, Q. Zhou and G.L. Benwell, “Using artificial neural networks to model the suitability of coastline for breeding by New Zealand fur seals (Arctocephalus forsteri )”, Ecological Modelling, vol. 148, 2002, pp. 111-131.

[10] Y. Hou and Y. Liu, Application of decision tree on land suitability analysis. In: D. Li, J. Gong, and H. Wu (Eds.), International Conference on Earth Observation Data Processing and Analysis (ICEODPA), Proceedings of SPIE, 2008, pp. 72854I.1-72854I.6.

[11] S. Kalogirou, “Expert systems and GIS: an application of land suitability evaluation”, Computers, Environment and Urban Systems, vol. 26, 2002, pp. 89-112.

[12] M.H. Tseng, S.J. Chen, G.H. Hwang and M.Y. Shen, “A genetic algorithm rule-based approach for land-cover classification”, ISPRS Journal of Photogrammetry and Remote Sensing, vol. 63 (2), 2008, pp. 202-212.

[13] J. Yu, Y. Chen and J. Wu, “Cellular Automata Based Spatial Multi-criteria Land Suitability Simulation for Irrigated Agriculture”, International Journal of Geographical Information Science. vol. 25(1), 2011, pp. 131-148.

[14] F. Hayes-Roth, “Rule-based systems”. Communications of the ACM, vol. 28, 1985, 921-932.

[15] T.R. Nisar Ahamed, K. Gopal Rao and J.S.R. Murthy, “GIS-based fuzzy membership model for crop-land suitability analysis”. Agricultural Systems, vol. 63, 2000, pp. 75-95.

[16] E. Van Broekhoven, V. Adriaenssens, B. De Baets and P.F.M. Verdonschot, “Fuzzy rule-based macroinvertebrate habitat suitability models for running waters”. ecological modeling, vol. 198, 2006, pp. 71-84.

[17] C. Blum and M. Dorigo, “Deception in ant colony optimization”. In: M. Dorigo, M. Birattari, C. Blum, L. M. Gambardella, F. Mondada, and T. Stutzle (Eds.), Proc. ANTS 2004, Fourth Internat. Workshop on Ant Colony Optimization and Swarm Intelligence, Lecture Notes in Computer Science, Berlin: Springer, 2004, pp. 119-130.

[18] M. Dorigo, G. Di Caro and L.M. Gambardella, “Ant algorithms for discrete optimization”, Artificial Life, vol. 5 (2), 1999, pp. 137-72.

[19] D. Merkle, M. Middendorf and H. Schmeck, “Ant colony optimization for resource-constrained project scheduling”, IEEE Trans Evolutionary Comput, vol. 6 (4), 2002, pp. 333-346.

[20] G. Fuellerer, K.F. Doerner, R.F. Hartl and M. Iori, “Ant colony optimization for the two-dimensional loading vehicle routing problem”, Computers and Operations Research, vol. 36 (3), 2009, pp. 655-673.

[21] R.S. Parpinelli, H.S. Lopes and A.A. Freitas, “Data mining with an ant colony optimization algorithm”. IEEE Transactions on Evolutionary Computation, vol. 6 (4), 2002, pp. 321-332.

[22] B. Liu, H.A. Abbass and B. Mckay, “Density-based heuristic for rule discovery with Ant-Miner”. In: The 6th Australia-Japan Joint Workshop on Intelligent and Evolutionary System, Canberra, Australia, 2002, pp. 180-184.

[23] Z. Wang and B. Feng, Classification Rule Mining with an Improved Ant Colony Algorithm. In: G. I. Webb, and X. Yu (Eds.), AI 2004: Advances in Artificial Intelligence, Lecture Notes in Computer Science, Berlin: Springer, 2005, pp. 357-367.

Page 6: [IEEE 2012 First International Conference on Agro-Geoinformatics - Shanghai, China (2012.08.2-2012.08.4)] 2012 First International Conference on Agro- Geoinformatics (Agro-Geoinformatics)

[24] G.A. Harris, “Soil Limitations of Irrigated Cropping on Macintyre Brook (Project Report Q086017)” (46pp), Brisbane: Queensland Department of Primary Industries, 1986.

[25] G.H. Malcolmson and P.L. Lloyd, “Inglewood Shire Handbook” (103pp), Brisbane: Queensland Department of Primary Industries, 1977.

[26] Y. Chen, S. Khan and Z. Paydar, “To Retire or Expand? A Fuzzy GIS-based Spatial Multi-criteria Evaluation Framework for Irrigated Agriculture”, Irrigation and Drainage, vol. 59(2), 2010, pp. 174-188.

[27] Z. Paydar, D. Gaydon and Y. Chen, “Up-scaling irrigation losses”, Irrigation Science, vol. 27, 2009, pp. 347-356.

[28] J. R. Quinlan, C4.5: Programs for machine learning. San Mateo, US: Morgan Kaufmann Publishers, 1993.