- Email: [email protected]
Development of an intelligent environmental knowledge system for sustainable agricultural decision support
Environmental Modelling & Software 52 (2014) 264e272
Contents lists available at ScienceDirect
Environmental Modelling & Software journal homepage: www.elsevier.com/locate/envsoft
Development of an intelligent environmental knowledge system for sustainable agricultural decision support Ritaban Dutta a, *, Ahsan Morshed a, Jagannath Aryal b, Claire D’Este a, Aruneema Das c a
Intelligent Sensing and Systems Laboratory, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Hobart 7001, Australia School of Geography and Environmental Studies, University of Tasmania, Hobart 7001, Australia c School of Engineering, University of Tasmania, Hobart 7001, Australia b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 9 May 2013 Received in revised form 1 October 2013 Accepted 3 October 2013 Available online 15 November 2013
The purpose of this research was to develop a knowledge recommendation architecture based on unsupervised machine learning and unified resource description framework (RDF) for integrated environmental sensory data sources. In developing this architecture, which is very useful for agricultural decision support systems, we considered web based large-scale dynamic data mining, contextual knowledge extraction, and integrated knowledge representation methods. Five different environmental data sources were considered to develop and test the proposed knowledge recommendation framework called Intelligent Environmental Knowledgebase (i-EKbase); including Bureau of Meteorology SILO, Australian Water Availability Project, Australian Soil Resource Information System, Australian National Cosmic Ray Soil Moisture Monitoring Facility, and NASA’s Moderate Resolution Imaging Spectroradiometer. Unsupervised clustering techniques based on Principal Component Analysis (PCA), Fuzzy-CMeans (FCM) and Self-organizing map (SOM) were used to create a 2D colour knowledge map representing the dynamics of the i-EKbase to provide “prior knowledge” about the integrated knowledgebase. Prior availability of recommendations from the knowledge base could potentially optimize the accessibility and usability issues related to big data sets and minimize the overall application costs. RDF representation has made i-EKbase flexible enough to publish and integrate on the Linked Open Data cloud. This newly developed system was evaluated as an expert agricultural decision support for sustainable water resource management case study in Australia at Tasmania with promising results. Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved.
Keywords: Feature Semantic matching Knowledge integration i-EKbase Linked Open Data cloud
Software and/or data availability NASA MODIS data is publicly available from NASA ASRIS is publicly available data base from CSIRO, Australia Jena framework for data translation http://jena.apache.org/ Sesame TripleStore http://www.openrdf.org/Pubby tool is Linked Open Data front end tool http://wifo5-03.informatik.uni-mannheim.de/pubby/
1. Problem space and motivation The ultimate challenge of an environmental and agricultural decision support system is to overcome uncertainty associated with
* Corresponding author. Tel.: þ61 3 6232 5423; fax: þ61 2 6232 5229. E-mail addresses: [email protected], [email protected] (R. Dutta).
the data quality, to cross validate the knowledge automatically, and to improve the efficiency of the decision making process. This research study proposed a knowledge integration platform (Knapen et al., 2013; Ren et al., 1988; Sheth, 2012; Henson et al., 2012) and machine learning analysis based recommendation architecture (Intelligent Environmental Knowledgebase (i-EKbase)) (Morshed et al., 2013a,b; Gilbert et al., 2010) to represent the knowledge in a more meaningful way. Knowledge integration and expert knowledge representation for sustainable agricultural decision support systems (SADSS) is an ever changing domain problem for environmental modelling and associated software system development. Often sensory system and model integrations do not reflect the natural fluidity (Voinov and Shugart, 2013) of the environment hence create significant performance limitations. In many cases SADSS also does not reflect or satisfy the end user requirements due to lack of domain knowledge capturing (McIntosh et al., 2011). These early works have motivated this research project to consider heterogeneous data sources and domain information systems as dynamic software modules before the knowledge integration to capture the natural dynamics of the environment in a more realistic
1364-8152/$ e see front matter Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.envsoft.2013.10.004
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
way possible. Automated multi-scale knowledge integration in a software system for emergency management has been reported (Akbar et al., 2013). It is also evident from literature that awareness and handling of uncertainty in SADSS could only be addressed by using multi-source multi-criteria knowledge integration and recommendation approach (Nino-Ruiz et al., 2013; Arnold, 2013). These previous reporting motivated the main focus of this paper to make the environmental knowledge more robust and trustworthy in a decision support context applicable to agricultural design. The application of i-EKbase to sustainable water resource management was studied in this paper as a proof of concept. In Australia, water usage for irrigation and associated electricity costs are extremely high and provides a compelling case study for the i-EKbase. Selection of this generic irrigation water management case study was inspired by the fact that there is a strong requirement for demonstrating the effectiveness of multi-source multi-scale spatiotemporal environmental knowledge integration and usefulness of
265
such approach for much wider performance improvement in various environmental decision support systems. Five different environmental data sources were considered for the development of i-EKbase, namely, SILO (SILO Website), AWAP (AWAP Website), CosmOz (COSMOZ Website), ASRIS (ASRIS Website), and MODIS (MODIS Website) (Fig. 1). The Long Paddock SILO database is operated by the Queensland Climate Change Centre of Excellence (QCCCE) from the Australian Department of Science, Information Technology, Innovation and the Arts (DSITIA). The Australian Water Availability Project (AWAP) database is developed to monitor the state and trend of the terrestrial water balance of the Australian continent, using modeldata fusion methods to combine both measurements and modelling. The Australian Soil Resource Information System (ASRIS) database provides online access to the best publicly available information on soil and land resources in a consistent format across Australia. The Australian Cosmic Ray Sensor Network (CosmOz)
Fig. 1. (a) Tullochgorum, Tasmania is represented on a MODIS image as the selected study location for this study, (b) ASRIS maps showing clay content, bulk density, and the plant available water capacity for Australia. A web adaptor was used to extract static data for a particular location, (c) Examples of AWAP gridded map data, (d) Example of SILO data time series extracted using dynamic web adaptor, (e) CosmOz Hydroinnova CRS-1000 cosmic ray soil moisture probe based neutron count data relating to area-average soil moisture over its horizontal footprint.
266
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
database is a near-real time soil moisture measurement network providing neutron counts related to bulk soil moisture. The MODIS (MODerate resolution Imaging Spectroradiometer) database which includes data from Terra and Aqua satellites e viewing the entire Earth’s surface every 1e2 days, acquiring data in 36 spectral bands, or groups is available via the NASA website. Individual web data adaptors were created to access and download data and integrate automatically based on semantic metadata matching mechanism. For ASRIS data processing we used ArcGIS 10.1 software provided by the Environmental Scientific Research Institute (ESRI), USA. For MODIS image data processing we used ENVI software provided by EXELIS. The i-EKbase has been developed as an automatically adaptable dynamic knowledgebase. Tullochgorum, Tasmania (Fig. 2) was selected as case study location for this paper which was described by latitude 41.7 and longitude 147.9. All data sets were downloaded and processed for the period 2010/12/15e2012/10/06 to develop a historic test version of i-EKbase dedicated for Tullochgorum (Morshed et al., 2013a,b). 2. Knowledge integration architecture Fig. 3 shows the Knowledge integration architecture proposed in this study. 2.1. Data adaptors layer Individual web data adaptors were created to access and download data automatically. Dynamic data availability made iEKbase an adaptive knowledgebase. The nearest Bureau of Meteorology (BOM) weather station to the location was selected based on distance for which the corresponding SILO data file was downloaded and processed. CosmOz Data was also downloaded for the selected BOM weather station. The AWAP database was
connected through a secured FTP server and grid files were downloaded locally. The ASRIS database was downloaded from the publicly available ASRIS website. For the same location a pixel position was derived on the daily continental AWAP gridded map and time series were extracted for individual variable for a given time frame. Similarly a pixel position was also calculated from the ASRIS data to extract soil resource information for the same latitude and longitude. MODIS images were downloaded and processed to extract time series data (Sheth 2012; Henson et al., 2012; Gilbert et al., 2010; Morshed et al., 2013a,b). 2.2. Data pre-processing layer The pre-processing of downloaded time series was an important feature due to the uncertainty associated with data availability. Individual time series were identified according to the name of the selected site and environmental variable. As can be expected in real world networks, each of the available time series had several periods with missing values. Two approaches were taken for handling the missing values in the pre-processing layer. Cubic interpolation was used to impute the missing values to fill up the gaps in a time series. On the other hand segments with missing values were considered as a unique feature state of the system. In the later approach instead of interpolating missing values, those segments were labelled as ‘Not a Number’. This was done to avoid introduction of artificial data uncertainty within the processing. Initially a filter was designed to remove all of the Infinite values, and replace them with a ‘Not a Number’ string to keep the filtering statistically insignificant and the original time frame unaltered. In the next stage of data pre-processing, context based filtering was applied. A sensor measuring a particular environmental parameter should operate within a well defined range. Hence, any value recorded outside of the operational range was treated as invalid data and replaced with a ‘Not a Number’ string. Filtered data was stored in an integrated structured array where columns represented different variables whereas rows represented time based observations (Miles and Bechhofer, 2009; Broekstra et al., 2002). In the i-EKbase instance for the Tullochgorum site had 40 environmental attributes in this response matrix (Fig. 4). Matrix had 40 columns and n number of rows depending on total time frame it represented. 2.3. Semantic matching layer Pre-processed data was cross validated using semantic matching and statistical cross correlation calculation. Based on natural language processing and sensor-model Ontologies a cross validation layer was created (Morshed et al., 2013a,b). Ideally, all similar environmental variables from different data sources should be able to cross validate each other statistically. Variables were semantically matched to form several subgroups, i.e. {SILO Max Temperature ( C), SILO Min Temperature ( C), AWAP Max Temperature ( C), AWAP Min Temperature ( C), MODIS/Terra Land Surface Temperature ( C)}. Based on semantic matching a pool of equivalent or related time series was created (Broekstra et al., 2002; W3C Website; Berners-Lee, 2011). 2.4. Complementing cross validation layer
Fig. 2. The map of the study area Tullochgorum in Tasmania.
The ‘cross-correlation technique’ was used to measure the similarities between two time series signals representing similar scenarios. If the two signals being compared were identical then the cross-correlation coefficient should be equal to 1, and if there are no similarities between the signals it should be equal to 0. A scoring protocol was designed on cross correlation results. The time series with highest score was selected from each sub group as best
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
267
Fig. 3. Knowledge integration architecture for the i-EKbase system.
Fig. 4. The details of these 40 environmental attributes integrated into the dynamic response matrix including definition of the variables and units (if applicable). For unified representation of the matrix all the variables were normalized within [0 to 1] range.
268
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
representative of the associated environmental variables from that pool (Morshed et al., 2013a,b). Missing values of a time series were complemented using other source based equivalent time series of the same pool.
users and application to interact with data, understanding and pattern matching from the data (Ren et al.,1988; Sheth, 2012; Henson et al., 2012; Gilbert et al., 2010; Morshed et al., 2013a,b; Broekstra et al., 2002; W3C Website; Berners-Lee, 2011; Bizer et al., 2009; Auer and Lehmann, 2010; Heath and Bizer, 2011; Mendes et al., 2012).
2.5. RDF conversion and integration layer 3. Knowledge recommendation The World Wide Web Consortium (W3C) introduced a format called Resource description framework (RDF) which is now a standard model for machine-readable data presentation. It decomposes data into the pieces (subject, object and predicate) and gives a uniform resource identifier (URI) for each resource or object. Through the URIs, it is possible to read the information about the particular resource on the web by using the HTTP access. The RDF has features that facilitate data integration even if the underlying schema differs, and it specially supports the evaluation of schemas over time without requiring the entire data consumer to be changed. Data from different sources were converted into RDF format and stored in a triplestore called Sesame. This triple store is a backbone of i-EKbase. Data were in .txt, .csv, .NetCDF, .HDF, and .shp format. Individual data wrappers were developed to convert them into unified RDF format (Fig. 5). The cross validated integrated data acquired for the Tullochgorum site was converted into RDF format (Miles and Bechhofer, 2009; Broekstra et al., 2002; W3C Website; Berners-Lee, 2011; Bizer et al., 2009; Auer and Lehmann, 2010; Heath and Bizer, 2011; Mendes et al., 2012). Fig. 3 also shows the RDF triples from different data sources. These triples are connected to each other. By using the triples, i-EKbase has the flexibility to be connected to the Linked Open Data cloud (LOD Cloud Website) and navigate through wider knowledge range information. 2.6. TripleStore and SPARQL queries implementation Integrated knowledge RDF files were uploaded into the triplestore which provided all facilities of browsing, querying, exporting data in different formats, etc. SPARQL (SPARQL Protocol and RDF Query Language) is query language for RDF data. SPARQL is a syntactically-SQL-like language for querying RDF graphs via pattern matching. The language’s features include basic conjunctive patterns, value filters, optional patterns, and pattern disjunction. SPAQRQL based i-EKbase Query was important development in the context of the semantic web, since it has provided a mechanism for
Prior availability of recommendation and knowledge on knowledgebase could potentially optimize the accessibility and usability issues related to large data sets. Further, it could minimize the overall computational and application costs. Unsupervised machine learning based visual knowledge recommendation was the main aspect of this system. 3.1. Hybrid unsupervised clustering Clustering algorithms based on Principal Component Analysis (PCA), Fuzzy-C-Means (FCM) were used to process the integrated data. The objective of this analysis was to establish a dynamic list of least correlated attributes that were contributing towards the most data variance. In PCA, each of the orthogonal components accounts for a certain amount of variance in the data, with a decreasing degree of importance. FCM clustering provided partitioning results with additional information supplied by the cluster membership values indicating the degrees of belongingness, where ‘C’ was the total number of clusters. Optimum number of clusters was determined by optimizing an objective function based on intra-cluster distance measures, which were 11 for this study. FCM was applied on the selected least correlated attributes from PCA to estimate the natural grouping of the data. This dynamic data analysis provided ranked attributes (according to their importance) which were considered as recommendations about the whole integrated data set for any future application design (Fig. 6). The final membership function outputs were used to design and initialize the SOM. The SOM algorithm was developed to transform an incoming signal pattern of arbitrary dimension into a one or two-dimensional discrete map. PCA and FCM analysis were used to guide the SOM clustering in terms of initializing the weights. It was evident from PCA that 5 attributes did not carry any data variance, so a SOM network of size [7 5 representing rest of the 35 variables] was created to train and visualize the natural grouping of the
Fig. 5. Example of RDF format created from the heterogeneous data sources.
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
269
Fig. 6. PCA-FCM-SOM based 2D visual knowledge recommendation system. Least correlated attributes would be marked with different colour patterns hence could easily be selected. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
variables. SOM clustering on the membership groups was quite useful as this technique provided a 2D visual colour map of the individual attribute’s trained SOM weights and characterized the natural grouping or similarity measures of that attribute with others. Less correlated attributes could be selected for an application based on quick visual inspection (Lin and Lee, 1996; Kohonen, 1982; Dutta and Dutta, 2006).
4. i-EKbase based water resource management In Australia, much of the environmental degradation including salinisation is associated with the changes in the near-surface water balance induced by massive clearing of native vegetation and deforestation. These artificial changes have led to significant increases in groundwater recharge, which in turn have led to rising
Fig. 7. Historic water balance indicator estimation based on integrated i-EKbase and PCA-FCM-SOM based attribute recommendations.
270
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
Fig. 8. Agricultural water requirement indicator and i-EKbase based estimated water balance profile for Tullochgorum location.
water tables and salinisation. Other important aspect of the hydrological system is vegetation. Vegetation influences the hydrological cycle through the exchange of energy, water, carbon and other substances and is therefore critical for many hydrological processes, in particular transpiration, infiltration and runoff. The movement of water through the hydrological cycle varies significantly in both time and space. Australia, the driest continent, has the highest variability in rainfall and runoff and is therefore a difficult system to model. i-EKbase was evaluated as an expert agricultural decision support system related to the catchment water balance model. Water balance is based on the law of conservation of mass: any change in the water content of a given soil volume during a specified period must equal the difference between the amount of water added to the soil volume and withdrawn from it (Zhang et al., 1996). When the control volume is the entire catchment represented by the given latitude and longitude information, the surface water balance equation can be expressed as:
DhSi ¼ hPi hETi hQ i hRi
(1)
Where DhSi is the change in spatially averaged catchment water storage, hPi is spatially averaged precipitation that goes to soil, hETi is the spatially averaged catchment evapotranspiration, hQ i is the spatially averaged catchment surface runoff, and hRi is the spatially averaged catchment recharge. This equation was used for i-EKbase based historic water balance calculation. As described in Fig. 7, an irrigation water requirement indicator was calculated using Equation (1). This indicator was calculated for the whole duration of this research experiment. Historically there were only two possible water management decisions that one farmer could take on a day e either they did require buying extra water for crop irrigation or there was enough water in soil that they did not require to buy extra water. Positive results from i-EKbase for water balance calculation indicated enough water in the soil (represented by Class 1 or ‘1’ values) where as negative results indicated the necessity of buying water (represented by Class 2 or ‘2’ values). A new time series was created to represent the variance of irrigation water requirement indicator over two complete years. Fig. 8 shows the
irrigation indicator profile for the Tullochgorum locations during the experimental period. 4.1. Supervised machine learning (SML) based generalization In a SML engine, a set of known samples (or known data) are systematically introduced to the learning algorithm, which updates associated weight vectors, and internally classifies data according to the known training targets or classes held in a knowledge base. In the second stage for identification, an unknown sample is tested against the knowledge base and then the membership class is predicted. Unknown samples are analysed using relationships found in the initial calibration, learning or training stage. Four supervised estimators, namely Back Propagation Multilayer Perceptron Network (BP-MLPN), Elman Networks (EN), Boltzmann Machine Network (BMN) and Radial Basis Function Network (RBFN) were trained and tested independently for this study to evaluate the performance of i-EKbase on agricultural water resource management decision support system. In this case 40 different environmental attributes of i-EKbase were training inputs (40 different time series representing daily data for same time period) with the irrigation indicator (one time series with same time length) based on water balance model was the training target (calculated from 17 selected variables as described in Fig. 7). Any new daily data combination including all 40 attributes would then be processed as testing input by the trained algorithm to predict the probable water balance situation as Class 1 or Class 2. That class prediction would then be considered and interpreted as vital water
Table 1 SML evaluation results for Tullochgorum.
Data set 1 Data set 2 Data set 3 Sensitivity Specificity
RBFN
BPO-MLPN
EN
BMN
75.5 91.3 82.1 90 85
62.5 71.7 76.1 79 63.3
73.9 70.5 72.9 65 76.5
68.1 78.5 81.5 61.5 69.2
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
271
Fig. 9. i-EKbase knowledge recommendation interface for Linked Open Data cloud.
management decision (Morshed et al., 2013a,b; Lin and Lee, 1996; Kohonen, 1982; Dutta and Dutta, 2006). 4.2. Experimental results Various training and testing data sets were formed based on a randomized incremental optimization algorithm to evaluate the generalization capability of the proposed individual and ensemble machine learning architectures. Combinations of % training data and % testing data were varied from {10%e90%} to {50%e50%} to identify the best possible training-testing data balance to achieve maximum prediction accuracy with highest possible sensitivity and specificity. Table 1 summarizes all the machine learning generalization results for the three different evaluation data sets, namely, DATA SET 1: {Training 90% e Testing10%}, DATA SET 2: {Training 70% e Testing 30%} and DATA SET 3: {Training 50% e Testing 50%}. RBFN was the best performer compared to the other three estimators while lesser amounts of data were used from training. Best result for Tullochgorum was 91.3% correct prediction using DATA SET 2 where sensitivity and specificity was 85% and 90% respectively. So achieved level of generalization was very encouraging in the context of predicting water balance and irrigation water requirement, using i-EKbase.
This issue has been addressed by using multi dimensional and heterogeneous data in i-EKbase system. The machine learning based data driven approach behind i-EKbase system was demonstrated as an effective method to estimate water balance with potentially higher accuracy. Supervised machine learning paradigms were experimented to explore generalization capability and prediction accuracy of this proposed water resource management solution based on multi sensor e model integration. The Radial Basis Function Network based 91.3% accuracy performance proved that newly proposed predictive water resource estimation method based on large multi scale knowledge integration could potentially make the irrigation decision support systems more robust and efficient. Acknowledgement The authors wish to thank the Intelligent Sensing and Systems Laboratory and the Tasmanian node of the Australian Centre for Broadband Innovation. Authors would like to thanks Andrew Terhorst, Peter Briggs, Rex Keen, Bill Cotching, Dave McJannet, Aaron Hawdon, Craig Lindley and Peter Taylor for their support of this work. References
5. i-EKbase published on Linked Open Data cloud i-EKbase was designed using URI and RDF to present as Linked Open Data as it is the best practice for exposing, sharing, and connecting pieces of data, information, and knowledge on the Semantic Web. A user interface layer (Fig. 9) was developed on top of these triples to provide flexibility to the application designers. The developed ontologies could be reused by others. In our lightweight ontologies, different standard vocabularies (i.e. SSN, SKOS) have been implemented for presenting our concepts. Furthermore, we used DBpedia and Geoname for connecting our resources to the LOD cloud. 6. Conclusion This paper has three main achievements. Firstly a multi-source environmental knowledge framework (i-EKbase) was developed to provide large-scale availability of relevant sensor-model databases for any environmental application. We have developed lightweight ontologies based on extracted metadata from heterogeneous data sources. Next historic surface water balance for one location in Tasmania, Australia was estimated using unsupervised machine learning knowledge recommendation. Currently, water balance is estimated using the conventional Equation (1) in Section 5. But this method does not include several other environmental parameters which might have significant influence on the water balance.
Akbar, M., Aliabadi, S., Patel, R., Watts, M., 2013. A fully automated and integrated multi-scale forecasting scheme for emergency preparedness. Environ. Model. Softw. 39, 24e38. Arnold, T.R., 2013. Procedural knowledge for integrated modelling: towards the modelling playground. Environ. Model. Softw. 26, 135e148. ASRIS Website http://www.asris.csiro.au/index_other.html (accessed in August 2013). Auer, S., Lehmann, J., 2010. Creating knowledge out of interlinked data. Semant. Web 1 (1), 97e104. AWAP Website http://www.eoc.csiro.au/awap/ (accessed in August 2013). Berners-Lee, T., 2011. Linked Data. http://www.w3.org/DesignIssues/LinkedData. html. Bizer, C., Heath, T., Berners-Lee, T., 2009. Linked data e the story so far. Int. J. Semant. Web Inf. Syst. 5 (3), 1e22. Broekstra, J., Kampman, A., van Harmelen, F., 2002. Sesame: a generic architecture for storing and querying RDF and RDF schema. In: The Semantic Web e ISWC 2002: First International Semantic Web Conference, 54e68, Sardinia, Italy. Springer, Berlin/Heidelberg. COSMOZ Website, 2013. http://www.ermt.csiro.au/html/cosmoz.html. Dutta, R., Dutta, R., 2006. “Maximum probability rule” based classification of MRSA infections in hospital environment: using electronic nose. Sens. Actuators B Chem. 120 (1), 156e165. Gilbert, R.C., Trafalis, T.B., Richman, M.B., Leslie, L.M., 2010. Machine learning methods for data assimilation. In: Computational Intelligence in Architecturing Complex Engineering Systems, pp. 105e112. New York, NY, USA. Heath, T., Bizer, C., 2011. Linked Data: Evolving the Web into a Global Data Space. In: Synthesis Lectures on the Semantic Web: Theory and Technology, vol. 1, pp. 1e 136. Henson, C.A., Sheth, A., Thirunarayan, P.K., 2012. Semantic perception: converting sensory observations to abstractions. IEEE Internet Comput. 16 (2), 26e34. Knapen, R., Janssen, S., Roosenschoon, O., Verweij, P., Winter, W., Uiterwijk, M., Wien, J., 2013. Evaluating OpenMI as a model integration platform across disciplines. Environ. Model. Softw. 39, 274e282.
272
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
Kohonen, T., 1982. Self-organized formation of topology correct feature maps. Biol. Cybern. 43, 59e69. Lin, C.T., Lee, C.S.G., 1996. Neural Fuzzy Systems a Neuro-Fuzzy Synergism to Intelligent Systems. Prentice Hall, Upper Saddle River, NJ 07458, ISBN 0-13235169-2. LOD Cloud Website http://lod-cloud.net/ (accessed in August 2013). McIntosh, B.S., Ascough, J.C., Twery, M., Chew, J., Elmahdi, A., Haase, D., Harou, J.J., Hepting, D., Cuddy, S., Jakeman, A.J., Chen, S., Kassahun, A., Lautenbach, S., Matthews, K., Merritt, W., Quinn, N.W.T., Rodriguez-Roda, I., Sieber, S., Stavenga, M., Sulis, A., Ticehurst, J., Volk, M., Wrobel, M., Delden, H., van ElSawah, S., Rizzoli, A., Voinov, A., 2011. Environmental decision support systems (EDSS) development e challenges and best practices. Environ. Model. Softw. 26, 1389e1402. Mendes, P., Mühleisen, H., Bizer, C., 2012. Sieve: linked data quality assessment and fusion. In: Invited Paper at the 1st International Workshop on Linked Web Data Management (LWDM 2012), Berlin, Germany. Miles, A., Bechhofer, S., 2009. W3C SKOS Simple Knowledge Organization System eXtension for Labels (SKOS-XL). http://www.w3.org/TR/skos-reference/skos-xl. html. MODIS Website http://modis.gsfc.nasa.gov/ (accessed in August 2013).
Morshed, A., Dutta, R., Aryal, J., 2013a. Recommending Environmental Knowledge as Linked Open Data Cloud Using Semantic Machine Learning. In: DESWeb 2013. ICDE, Brisbane, Australia, pp. 27e28. Morshed, A., Aryal, J., Dutta, R., 2013b. Environmental spatio-temporal ontology for the Linked Open Data cloud. In: 12th IEEE TrustCom 2013 (iWCDM), Melbourne, Australia. Nino-Ruiz, M., Bishop, I., Pettit, C., 2013. Spatial model steering, an exploratory approach to uncertainty awareness in land use allocation. Environ. Model. Softw. 39, 70e80. Ren, C.L., Lin, M.-H., Scherp, R.S., 1988. Dynamic multi-sensor data fusion system for intelligent robots. IEEE J. Robot. Autom. 4, 386e396. Sheth, A.P., 2012. A new landscape for distributed and parallel data management. Distrib. Parallel Databases 30 (2), 101e103. SILO Website http://www.longpaddock.qld.gov.au/silo/ (accessed in August 2013). Voinov, A., Shugart, H., 2013. ‘Integronsters’, integral and integrated modeling. Environ. Model. Softw. 39, 149e158. W3C Website. Resource Describe Framework. Retrieve at: http://www.w3.org/RDF/ (accessed in August 2013). Zhang, L., Dawes, W.R., Hatton, T.J., 1996. Modelling hydrological processes using a biophysically based modeldapplication of WAVES to FIFE and HAPEXMOBILHY. J. Hydrol. 185, 147e169.
Contents lists available at ScienceDirect
Environmental Modelling & Software journal homepage: www.elsevier.com/locate/envsoft
Development of an intelligent environmental knowledge system for sustainable agricultural decision support Ritaban Dutta a, *, Ahsan Morshed a, Jagannath Aryal b, Claire D’Este a, Aruneema Das c a
Intelligent Sensing and Systems Laboratory, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Hobart 7001, Australia School of Geography and Environmental Studies, University of Tasmania, Hobart 7001, Australia c School of Engineering, University of Tasmania, Hobart 7001, Australia b
a r t i c l e i n f o
a b s t r a c t
Article history: Received 9 May 2013 Received in revised form 1 October 2013 Accepted 3 October 2013 Available online 15 November 2013
The purpose of this research was to develop a knowledge recommendation architecture based on unsupervised machine learning and unified resource description framework (RDF) for integrated environmental sensory data sources. In developing this architecture, which is very useful for agricultural decision support systems, we considered web based large-scale dynamic data mining, contextual knowledge extraction, and integrated knowledge representation methods. Five different environmental data sources were considered to develop and test the proposed knowledge recommendation framework called Intelligent Environmental Knowledgebase (i-EKbase); including Bureau of Meteorology SILO, Australian Water Availability Project, Australian Soil Resource Information System, Australian National Cosmic Ray Soil Moisture Monitoring Facility, and NASA’s Moderate Resolution Imaging Spectroradiometer. Unsupervised clustering techniques based on Principal Component Analysis (PCA), Fuzzy-CMeans (FCM) and Self-organizing map (SOM) were used to create a 2D colour knowledge map representing the dynamics of the i-EKbase to provide “prior knowledge” about the integrated knowledgebase. Prior availability of recommendations from the knowledge base could potentially optimize the accessibility and usability issues related to big data sets and minimize the overall application costs. RDF representation has made i-EKbase flexible enough to publish and integrate on the Linked Open Data cloud. This newly developed system was evaluated as an expert agricultural decision support for sustainable water resource management case study in Australia at Tasmania with promising results. Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved.
Keywords: Feature Semantic matching Knowledge integration i-EKbase Linked Open Data cloud
Software and/or data availability NASA MODIS data is publicly available from NASA ASRIS is publicly available data base from CSIRO, Australia Jena framework for data translation http://jena.apache.org/ Sesame TripleStore http://www.openrdf.org/Pubby tool is Linked Open Data front end tool http://wifo5-03.informatik.uni-mannheim.de/pubby/
1. Problem space and motivation The ultimate challenge of an environmental and agricultural decision support system is to overcome uncertainty associated with
* Corresponding author. Tel.: þ61 3 6232 5423; fax: þ61 2 6232 5229. E-mail addresses: [email protected], [email protected] (R. Dutta).
the data quality, to cross validate the knowledge automatically, and to improve the efficiency of the decision making process. This research study proposed a knowledge integration platform (Knapen et al., 2013; Ren et al., 1988; Sheth, 2012; Henson et al., 2012) and machine learning analysis based recommendation architecture (Intelligent Environmental Knowledgebase (i-EKbase)) (Morshed et al., 2013a,b; Gilbert et al., 2010) to represent the knowledge in a more meaningful way. Knowledge integration and expert knowledge representation for sustainable agricultural decision support systems (SADSS) is an ever changing domain problem for environmental modelling and associated software system development. Often sensory system and model integrations do not reflect the natural fluidity (Voinov and Shugart, 2013) of the environment hence create significant performance limitations. In many cases SADSS also does not reflect or satisfy the end user requirements due to lack of domain knowledge capturing (McIntosh et al., 2011). These early works have motivated this research project to consider heterogeneous data sources and domain information systems as dynamic software modules before the knowledge integration to capture the natural dynamics of the environment in a more realistic
1364-8152/$ e see front matter Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.envsoft.2013.10.004
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
way possible. Automated multi-scale knowledge integration in a software system for emergency management has been reported (Akbar et al., 2013). It is also evident from literature that awareness and handling of uncertainty in SADSS could only be addressed by using multi-source multi-criteria knowledge integration and recommendation approach (Nino-Ruiz et al., 2013; Arnold, 2013). These previous reporting motivated the main focus of this paper to make the environmental knowledge more robust and trustworthy in a decision support context applicable to agricultural design. The application of i-EKbase to sustainable water resource management was studied in this paper as a proof of concept. In Australia, water usage for irrigation and associated electricity costs are extremely high and provides a compelling case study for the i-EKbase. Selection of this generic irrigation water management case study was inspired by the fact that there is a strong requirement for demonstrating the effectiveness of multi-source multi-scale spatiotemporal environmental knowledge integration and usefulness of
265
such approach for much wider performance improvement in various environmental decision support systems. Five different environmental data sources were considered for the development of i-EKbase, namely, SILO (SILO Website), AWAP (AWAP Website), CosmOz (COSMOZ Website), ASRIS (ASRIS Website), and MODIS (MODIS Website) (Fig. 1). The Long Paddock SILO database is operated by the Queensland Climate Change Centre of Excellence (QCCCE) from the Australian Department of Science, Information Technology, Innovation and the Arts (DSITIA). The Australian Water Availability Project (AWAP) database is developed to monitor the state and trend of the terrestrial water balance of the Australian continent, using modeldata fusion methods to combine both measurements and modelling. The Australian Soil Resource Information System (ASRIS) database provides online access to the best publicly available information on soil and land resources in a consistent format across Australia. The Australian Cosmic Ray Sensor Network (CosmOz)
Fig. 1. (a) Tullochgorum, Tasmania is represented on a MODIS image as the selected study location for this study, (b) ASRIS maps showing clay content, bulk density, and the plant available water capacity for Australia. A web adaptor was used to extract static data for a particular location, (c) Examples of AWAP gridded map data, (d) Example of SILO data time series extracted using dynamic web adaptor, (e) CosmOz Hydroinnova CRS-1000 cosmic ray soil moisture probe based neutron count data relating to area-average soil moisture over its horizontal footprint.
266
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
database is a near-real time soil moisture measurement network providing neutron counts related to bulk soil moisture. The MODIS (MODerate resolution Imaging Spectroradiometer) database which includes data from Terra and Aqua satellites e viewing the entire Earth’s surface every 1e2 days, acquiring data in 36 spectral bands, or groups is available via the NASA website. Individual web data adaptors were created to access and download data and integrate automatically based on semantic metadata matching mechanism. For ASRIS data processing we used ArcGIS 10.1 software provided by the Environmental Scientific Research Institute (ESRI), USA. For MODIS image data processing we used ENVI software provided by EXELIS. The i-EKbase has been developed as an automatically adaptable dynamic knowledgebase. Tullochgorum, Tasmania (Fig. 2) was selected as case study location for this paper which was described by latitude 41.7 and longitude 147.9. All data sets were downloaded and processed for the period 2010/12/15e2012/10/06 to develop a historic test version of i-EKbase dedicated for Tullochgorum (Morshed et al., 2013a,b). 2. Knowledge integration architecture Fig. 3 shows the Knowledge integration architecture proposed in this study. 2.1. Data adaptors layer Individual web data adaptors were created to access and download data automatically. Dynamic data availability made iEKbase an adaptive knowledgebase. The nearest Bureau of Meteorology (BOM) weather station to the location was selected based on distance for which the corresponding SILO data file was downloaded and processed. CosmOz Data was also downloaded for the selected BOM weather station. The AWAP database was
connected through a secured FTP server and grid files were downloaded locally. The ASRIS database was downloaded from the publicly available ASRIS website. For the same location a pixel position was derived on the daily continental AWAP gridded map and time series were extracted for individual variable for a given time frame. Similarly a pixel position was also calculated from the ASRIS data to extract soil resource information for the same latitude and longitude. MODIS images were downloaded and processed to extract time series data (Sheth 2012; Henson et al., 2012; Gilbert et al., 2010; Morshed et al., 2013a,b). 2.2. Data pre-processing layer The pre-processing of downloaded time series was an important feature due to the uncertainty associated with data availability. Individual time series were identified according to the name of the selected site and environmental variable. As can be expected in real world networks, each of the available time series had several periods with missing values. Two approaches were taken for handling the missing values in the pre-processing layer. Cubic interpolation was used to impute the missing values to fill up the gaps in a time series. On the other hand segments with missing values were considered as a unique feature state of the system. In the later approach instead of interpolating missing values, those segments were labelled as ‘Not a Number’. This was done to avoid introduction of artificial data uncertainty within the processing. Initially a filter was designed to remove all of the Infinite values, and replace them with a ‘Not a Number’ string to keep the filtering statistically insignificant and the original time frame unaltered. In the next stage of data pre-processing, context based filtering was applied. A sensor measuring a particular environmental parameter should operate within a well defined range. Hence, any value recorded outside of the operational range was treated as invalid data and replaced with a ‘Not a Number’ string. Filtered data was stored in an integrated structured array where columns represented different variables whereas rows represented time based observations (Miles and Bechhofer, 2009; Broekstra et al., 2002). In the i-EKbase instance for the Tullochgorum site had 40 environmental attributes in this response matrix (Fig. 4). Matrix had 40 columns and n number of rows depending on total time frame it represented. 2.3. Semantic matching layer Pre-processed data was cross validated using semantic matching and statistical cross correlation calculation. Based on natural language processing and sensor-model Ontologies a cross validation layer was created (Morshed et al., 2013a,b). Ideally, all similar environmental variables from different data sources should be able to cross validate each other statistically. Variables were semantically matched to form several subgroups, i.e. {SILO Max Temperature ( C), SILO Min Temperature ( C), AWAP Max Temperature ( C), AWAP Min Temperature ( C), MODIS/Terra Land Surface Temperature ( C)}. Based on semantic matching a pool of equivalent or related time series was created (Broekstra et al., 2002; W3C Website; Berners-Lee, 2011). 2.4. Complementing cross validation layer
Fig. 2. The map of the study area Tullochgorum in Tasmania.
The ‘cross-correlation technique’ was used to measure the similarities between two time series signals representing similar scenarios. If the two signals being compared were identical then the cross-correlation coefficient should be equal to 1, and if there are no similarities between the signals it should be equal to 0. A scoring protocol was designed on cross correlation results. The time series with highest score was selected from each sub group as best
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
267
Fig. 3. Knowledge integration architecture for the i-EKbase system.
Fig. 4. The details of these 40 environmental attributes integrated into the dynamic response matrix including definition of the variables and units (if applicable). For unified representation of the matrix all the variables were normalized within [0 to 1] range.
268
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
representative of the associated environmental variables from that pool (Morshed et al., 2013a,b). Missing values of a time series were complemented using other source based equivalent time series of the same pool.
users and application to interact with data, understanding and pattern matching from the data (Ren et al.,1988; Sheth, 2012; Henson et al., 2012; Gilbert et al., 2010; Morshed et al., 2013a,b; Broekstra et al., 2002; W3C Website; Berners-Lee, 2011; Bizer et al., 2009; Auer and Lehmann, 2010; Heath and Bizer, 2011; Mendes et al., 2012).
2.5. RDF conversion and integration layer 3. Knowledge recommendation The World Wide Web Consortium (W3C) introduced a format called Resource description framework (RDF) which is now a standard model for machine-readable data presentation. It decomposes data into the pieces (subject, object and predicate) and gives a uniform resource identifier (URI) for each resource or object. Through the URIs, it is possible to read the information about the particular resource on the web by using the HTTP access. The RDF has features that facilitate data integration even if the underlying schema differs, and it specially supports the evaluation of schemas over time without requiring the entire data consumer to be changed. Data from different sources were converted into RDF format and stored in a triplestore called Sesame. This triple store is a backbone of i-EKbase. Data were in .txt, .csv, .NetCDF, .HDF, and .shp format. Individual data wrappers were developed to convert them into unified RDF format (Fig. 5). The cross validated integrated data acquired for the Tullochgorum site was converted into RDF format (Miles and Bechhofer, 2009; Broekstra et al., 2002; W3C Website; Berners-Lee, 2011; Bizer et al., 2009; Auer and Lehmann, 2010; Heath and Bizer, 2011; Mendes et al., 2012). Fig. 3 also shows the RDF triples from different data sources. These triples are connected to each other. By using the triples, i-EKbase has the flexibility to be connected to the Linked Open Data cloud (LOD Cloud Website) and navigate through wider knowledge range information. 2.6. TripleStore and SPARQL queries implementation Integrated knowledge RDF files were uploaded into the triplestore which provided all facilities of browsing, querying, exporting data in different formats, etc. SPARQL (SPARQL Protocol and RDF Query Language) is query language for RDF data. SPARQL is a syntactically-SQL-like language for querying RDF graphs via pattern matching. The language’s features include basic conjunctive patterns, value filters, optional patterns, and pattern disjunction. SPAQRQL based i-EKbase Query was important development in the context of the semantic web, since it has provided a mechanism for
Prior availability of recommendation and knowledge on knowledgebase could potentially optimize the accessibility and usability issues related to large data sets. Further, it could minimize the overall computational and application costs. Unsupervised machine learning based visual knowledge recommendation was the main aspect of this system. 3.1. Hybrid unsupervised clustering Clustering algorithms based on Principal Component Analysis (PCA), Fuzzy-C-Means (FCM) were used to process the integrated data. The objective of this analysis was to establish a dynamic list of least correlated attributes that were contributing towards the most data variance. In PCA, each of the orthogonal components accounts for a certain amount of variance in the data, with a decreasing degree of importance. FCM clustering provided partitioning results with additional information supplied by the cluster membership values indicating the degrees of belongingness, where ‘C’ was the total number of clusters. Optimum number of clusters was determined by optimizing an objective function based on intra-cluster distance measures, which were 11 for this study. FCM was applied on the selected least correlated attributes from PCA to estimate the natural grouping of the data. This dynamic data analysis provided ranked attributes (according to their importance) which were considered as recommendations about the whole integrated data set for any future application design (Fig. 6). The final membership function outputs were used to design and initialize the SOM. The SOM algorithm was developed to transform an incoming signal pattern of arbitrary dimension into a one or two-dimensional discrete map. PCA and FCM analysis were used to guide the SOM clustering in terms of initializing the weights. It was evident from PCA that 5 attributes did not carry any data variance, so a SOM network of size [7 5 representing rest of the 35 variables] was created to train and visualize the natural grouping of the
Fig. 5. Example of RDF format created from the heterogeneous data sources.
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
269
Fig. 6. PCA-FCM-SOM based 2D visual knowledge recommendation system. Least correlated attributes would be marked with different colour patterns hence could easily be selected. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
variables. SOM clustering on the membership groups was quite useful as this technique provided a 2D visual colour map of the individual attribute’s trained SOM weights and characterized the natural grouping or similarity measures of that attribute with others. Less correlated attributes could be selected for an application based on quick visual inspection (Lin and Lee, 1996; Kohonen, 1982; Dutta and Dutta, 2006).
4. i-EKbase based water resource management In Australia, much of the environmental degradation including salinisation is associated with the changes in the near-surface water balance induced by massive clearing of native vegetation and deforestation. These artificial changes have led to significant increases in groundwater recharge, which in turn have led to rising
Fig. 7. Historic water balance indicator estimation based on integrated i-EKbase and PCA-FCM-SOM based attribute recommendations.
270
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
Fig. 8. Agricultural water requirement indicator and i-EKbase based estimated water balance profile for Tullochgorum location.
water tables and salinisation. Other important aspect of the hydrological system is vegetation. Vegetation influences the hydrological cycle through the exchange of energy, water, carbon and other substances and is therefore critical for many hydrological processes, in particular transpiration, infiltration and runoff. The movement of water through the hydrological cycle varies significantly in both time and space. Australia, the driest continent, has the highest variability in rainfall and runoff and is therefore a difficult system to model. i-EKbase was evaluated as an expert agricultural decision support system related to the catchment water balance model. Water balance is based on the law of conservation of mass: any change in the water content of a given soil volume during a specified period must equal the difference between the amount of water added to the soil volume and withdrawn from it (Zhang et al., 1996). When the control volume is the entire catchment represented by the given latitude and longitude information, the surface water balance equation can be expressed as:
DhSi ¼ hPi hETi hQ i hRi
(1)
Where DhSi is the change in spatially averaged catchment water storage, hPi is spatially averaged precipitation that goes to soil, hETi is the spatially averaged catchment evapotranspiration, hQ i is the spatially averaged catchment surface runoff, and hRi is the spatially averaged catchment recharge. This equation was used for i-EKbase based historic water balance calculation. As described in Fig. 7, an irrigation water requirement indicator was calculated using Equation (1). This indicator was calculated for the whole duration of this research experiment. Historically there were only two possible water management decisions that one farmer could take on a day e either they did require buying extra water for crop irrigation or there was enough water in soil that they did not require to buy extra water. Positive results from i-EKbase for water balance calculation indicated enough water in the soil (represented by Class 1 or ‘1’ values) where as negative results indicated the necessity of buying water (represented by Class 2 or ‘2’ values). A new time series was created to represent the variance of irrigation water requirement indicator over two complete years. Fig. 8 shows the
irrigation indicator profile for the Tullochgorum locations during the experimental period. 4.1. Supervised machine learning (SML) based generalization In a SML engine, a set of known samples (or known data) are systematically introduced to the learning algorithm, which updates associated weight vectors, and internally classifies data according to the known training targets or classes held in a knowledge base. In the second stage for identification, an unknown sample is tested against the knowledge base and then the membership class is predicted. Unknown samples are analysed using relationships found in the initial calibration, learning or training stage. Four supervised estimators, namely Back Propagation Multilayer Perceptron Network (BP-MLPN), Elman Networks (EN), Boltzmann Machine Network (BMN) and Radial Basis Function Network (RBFN) were trained and tested independently for this study to evaluate the performance of i-EKbase on agricultural water resource management decision support system. In this case 40 different environmental attributes of i-EKbase were training inputs (40 different time series representing daily data for same time period) with the irrigation indicator (one time series with same time length) based on water balance model was the training target (calculated from 17 selected variables as described in Fig. 7). Any new daily data combination including all 40 attributes would then be processed as testing input by the trained algorithm to predict the probable water balance situation as Class 1 or Class 2. That class prediction would then be considered and interpreted as vital water
Table 1 SML evaluation results for Tullochgorum.
Data set 1 Data set 2 Data set 3 Sensitivity Specificity
RBFN
BPO-MLPN
EN
BMN
75.5 91.3 82.1 90 85
62.5 71.7 76.1 79 63.3
73.9 70.5 72.9 65 76.5
68.1 78.5 81.5 61.5 69.2
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
271
Fig. 9. i-EKbase knowledge recommendation interface for Linked Open Data cloud.
management decision (Morshed et al., 2013a,b; Lin and Lee, 1996; Kohonen, 1982; Dutta and Dutta, 2006). 4.2. Experimental results Various training and testing data sets were formed based on a randomized incremental optimization algorithm to evaluate the generalization capability of the proposed individual and ensemble machine learning architectures. Combinations of % training data and % testing data were varied from {10%e90%} to {50%e50%} to identify the best possible training-testing data balance to achieve maximum prediction accuracy with highest possible sensitivity and specificity. Table 1 summarizes all the machine learning generalization results for the three different evaluation data sets, namely, DATA SET 1: {Training 90% e Testing10%}, DATA SET 2: {Training 70% e Testing 30%} and DATA SET 3: {Training 50% e Testing 50%}. RBFN was the best performer compared to the other three estimators while lesser amounts of data were used from training. Best result for Tullochgorum was 91.3% correct prediction using DATA SET 2 where sensitivity and specificity was 85% and 90% respectively. So achieved level of generalization was very encouraging in the context of predicting water balance and irrigation water requirement, using i-EKbase.
This issue has been addressed by using multi dimensional and heterogeneous data in i-EKbase system. The machine learning based data driven approach behind i-EKbase system was demonstrated as an effective method to estimate water balance with potentially higher accuracy. Supervised machine learning paradigms were experimented to explore generalization capability and prediction accuracy of this proposed water resource management solution based on multi sensor e model integration. The Radial Basis Function Network based 91.3% accuracy performance proved that newly proposed predictive water resource estimation method based on large multi scale knowledge integration could potentially make the irrigation decision support systems more robust and efficient. Acknowledgement The authors wish to thank the Intelligent Sensing and Systems Laboratory and the Tasmanian node of the Australian Centre for Broadband Innovation. Authors would like to thanks Andrew Terhorst, Peter Briggs, Rex Keen, Bill Cotching, Dave McJannet, Aaron Hawdon, Craig Lindley and Peter Taylor for their support of this work. References
5. i-EKbase published on Linked Open Data cloud i-EKbase was designed using URI and RDF to present as Linked Open Data as it is the best practice for exposing, sharing, and connecting pieces of data, information, and knowledge on the Semantic Web. A user interface layer (Fig. 9) was developed on top of these triples to provide flexibility to the application designers. The developed ontologies could be reused by others. In our lightweight ontologies, different standard vocabularies (i.e. SSN, SKOS) have been implemented for presenting our concepts. Furthermore, we used DBpedia and Geoname for connecting our resources to the LOD cloud. 6. Conclusion This paper has three main achievements. Firstly a multi-source environmental knowledge framework (i-EKbase) was developed to provide large-scale availability of relevant sensor-model databases for any environmental application. We have developed lightweight ontologies based on extracted metadata from heterogeneous data sources. Next historic surface water balance for one location in Tasmania, Australia was estimated using unsupervised machine learning knowledge recommendation. Currently, water balance is estimated using the conventional Equation (1) in Section 5. But this method does not include several other environmental parameters which might have significant influence on the water balance.
Akbar, M., Aliabadi, S., Patel, R., Watts, M., 2013. A fully automated and integrated multi-scale forecasting scheme for emergency preparedness. Environ. Model. Softw. 39, 24e38. Arnold, T.R., 2013. Procedural knowledge for integrated modelling: towards the modelling playground. Environ. Model. Softw. 26, 135e148. ASRIS Website http://www.asris.csiro.au/index_other.html (accessed in August 2013). Auer, S., Lehmann, J., 2010. Creating knowledge out of interlinked data. Semant. Web 1 (1), 97e104. AWAP Website http://www.eoc.csiro.au/awap/ (accessed in August 2013). Berners-Lee, T., 2011. Linked Data. http://www.w3.org/DesignIssues/LinkedData. html. Bizer, C., Heath, T., Berners-Lee, T., 2009. Linked data e the story so far. Int. J. Semant. Web Inf. Syst. 5 (3), 1e22. Broekstra, J., Kampman, A., van Harmelen, F., 2002. Sesame: a generic architecture for storing and querying RDF and RDF schema. In: The Semantic Web e ISWC 2002: First International Semantic Web Conference, 54e68, Sardinia, Italy. Springer, Berlin/Heidelberg. COSMOZ Website, 2013. http://www.ermt.csiro.au/html/cosmoz.html. Dutta, R., Dutta, R., 2006. “Maximum probability rule” based classification of MRSA infections in hospital environment: using electronic nose. Sens. Actuators B Chem. 120 (1), 156e165. Gilbert, R.C., Trafalis, T.B., Richman, M.B., Leslie, L.M., 2010. Machine learning methods for data assimilation. In: Computational Intelligence in Architecturing Complex Engineering Systems, pp. 105e112. New York, NY, USA. Heath, T., Bizer, C., 2011. Linked Data: Evolving the Web into a Global Data Space. In: Synthesis Lectures on the Semantic Web: Theory and Technology, vol. 1, pp. 1e 136. Henson, C.A., Sheth, A., Thirunarayan, P.K., 2012. Semantic perception: converting sensory observations to abstractions. IEEE Internet Comput. 16 (2), 26e34. Knapen, R., Janssen, S., Roosenschoon, O., Verweij, P., Winter, W., Uiterwijk, M., Wien, J., 2013. Evaluating OpenMI as a model integration platform across disciplines. Environ. Model. Softw. 39, 274e282.
272
R. Dutta et al. / Environmental Modelling & Software 52 (2014) 264e272
Kohonen, T., 1982. Self-organized formation of topology correct feature maps. Biol. Cybern. 43, 59e69. Lin, C.T., Lee, C.S.G., 1996. Neural Fuzzy Systems a Neuro-Fuzzy Synergism to Intelligent Systems. Prentice Hall, Upper Saddle River, NJ 07458, ISBN 0-13235169-2. LOD Cloud Website http://lod-cloud.net/ (accessed in August 2013). McIntosh, B.S., Ascough, J.C., Twery, M., Chew, J., Elmahdi, A., Haase, D., Harou, J.J., Hepting, D., Cuddy, S., Jakeman, A.J., Chen, S., Kassahun, A., Lautenbach, S., Matthews, K., Merritt, W., Quinn, N.W.T., Rodriguez-Roda, I., Sieber, S., Stavenga, M., Sulis, A., Ticehurst, J., Volk, M., Wrobel, M., Delden, H., van ElSawah, S., Rizzoli, A., Voinov, A., 2011. Environmental decision support systems (EDSS) development e challenges and best practices. Environ. Model. Softw. 26, 1389e1402. Mendes, P., Mühleisen, H., Bizer, C., 2012. Sieve: linked data quality assessment and fusion. In: Invited Paper at the 1st International Workshop on Linked Web Data Management (LWDM 2012), Berlin, Germany. Miles, A., Bechhofer, S., 2009. W3C SKOS Simple Knowledge Organization System eXtension for Labels (SKOS-XL). http://www.w3.org/TR/skos-reference/skos-xl. html. MODIS Website http://modis.gsfc.nasa.gov/ (accessed in August 2013).
Morshed, A., Dutta, R., Aryal, J., 2013a. Recommending Environmental Knowledge as Linked Open Data Cloud Using Semantic Machine Learning. In: DESWeb 2013. ICDE, Brisbane, Australia, pp. 27e28. Morshed, A., Aryal, J., Dutta, R., 2013b. Environmental spatio-temporal ontology for the Linked Open Data cloud. In: 12th IEEE TrustCom 2013 (iWCDM), Melbourne, Australia. Nino-Ruiz, M., Bishop, I., Pettit, C., 2013. Spatial model steering, an exploratory approach to uncertainty awareness in land use allocation. Environ. Model. Softw. 39, 70e80. Ren, C.L., Lin, M.-H., Scherp, R.S., 1988. Dynamic multi-sensor data fusion system for intelligent robots. IEEE J. Robot. Autom. 4, 386e396. Sheth, A.P., 2012. A new landscape for distributed and parallel data management. Distrib. Parallel Databases 30 (2), 101e103. SILO Website http://www.longpaddock.qld.gov.au/silo/ (accessed in August 2013). Voinov, A., Shugart, H., 2013. ‘Integronsters’, integral and integrated modeling. Environ. Model. Softw. 39, 149e158. W3C Website. Resource Describe Framework. Retrieve at: http://www.w3.org/RDF/ (accessed in August 2013). Zhang, L., Dawes, W.R., Hatton, T.J., 1996. Modelling hydrological processes using a biophysically based modeldapplication of WAVES to FIFE and HAPEXMOBILHY. J. Hydrol. 185, 147e169.