XUE Cunjin, DONG Qing, and QIN Lijuan

Key Laboratory of Digital Earth Science,Institute of Remote Sensing and Digital Earth,Chinese Academy of Sciences,Beijing100094,P. R. China

A Cluster-Based Method for Marine Sensitive Object Extraction and Representation

XUE Cunjin*, DONG Qing, and QIN Lijuan

Key Laboratory of Digital Earth Science,Institute of Remote Sensing and Digital Earth,Chinese Academy of Sciences,Beijing100094,P. R. China

Within the context of global change, marine sensitive factors or Marine Essential Climate Variables have been defined by many projects, and their sensitive spatial regions and time phases play significant roles in regional sea-air interactions and better understanding of their dynamic process. In this paper, we propose a cluster-based method for marine sensitive region extraction and representation. This method includes a kernel expansion algorithm for extracting marine sensitive regions, and a field-object triple form, integration of object-oriented and field-based model, for representing marine sensitive objects. Firstly, this method recognizes ENSO-related spatial patterns using empirical orthogonal decomposition of long term marine sensitive factors and correlation analysis with multiple ENSO index. The cluster kernel, defined by statistics of spatial patterns, is initialized to carry out spatial expansion and cluster mergence with spatial neighborhoods recursively, then all the related lattices with similar behavior are merged into marine sensitive regions. After this, the Field-object triple form of ＜ O, A, F ＞ is used to represent the marine sensitive objects, both with the discrete object with a precise extend and boundary, and the continuous field with variations dependent on spatial locations. Finally, the marine sensitive objects about sea surface temperature are extracted, represented and analyzed as a case of study, which proves the effectiveness and the efficiency of the proposed method.

marine sensitive object; kernel-based expansion; Field-object model; remote sensing images; global change

1 Introduction

Advanced Earth-observing technologies make it possible to acquire long time series of marine bio-optical parameters, and dynamic parameters from multiple remote sensing images. Among these marine environmental parameters, some are sensitive to global changes, others may be not. On the condition of climate change, Sensitive Factors to Global Changes (SFGCs) have been proposed by Guo (2009), which include performance factors, response factors and driving factors. In the following years, the similar concepts, Essential Climate Variables (ECVs) have also been defined by Global Climate Observing System (GCOS), and marine ECVs include sea surface temperature, ocean color, sea level and sea ice (GCOS, 2011). These SFGCs and ECVs aim at characterizing the state of the global climate system and enabling long-term climate monitoring (Hollmannet al., 2013). However, these SFGCs are sensitive neither on global scale nor in all time series,i.e., in some regions and at some time phases, they are sensitive to climate changes, in some others they are not, such as the well-known El Ni?o re-gions, western Pacific warm pool, eastern Pacific cold tongue, ocean province (Esaiaset al., 2000),etc. And the spatiotemporal variations in different space and time play a more important role in better understanding their dynamic process on the condition of global climate changes (Oliver and Irwin, 2008; Milneet al., 2009).

How to extract the sensitive space regions and time phases from the multiple long term remote sensing images and to represent them underlying the interactions between sensitive factors and climate changes are two cutting-edge issues in scientific research.

Empirical orthogonal function (EOF) is an efficient method for extracting the dominant spatial patterns and associated temporal components by decomposing a spacetime field into orthogonal space and time functions (Hannachiet al., 2007), and is widely used in physical oceanography recently,i.e., spatiotemporal variation of sea surface temperature (SST) (Moron, 2003; Nardelliet al., 2010), spatial distribution and annual, seasonal and monthly properties of sea surface chlorophyll-aconcentration and marine primitive productivity (Shenet al., 2008; Zhan, 2008), and sea surface level anomaly (Chenet al., 2010). The dominant spatial pattern is to reveal the main spatial distribution with the similar variations in corresponding time components, so the uniform changing regions within spatial patterns are candidates for sensitiveregions.

The spatial pattern derived from remote sensing images is formatted with raster model, and has similarity-based groupings of values. Meanwhile, clusters are to represent groups of objects with similar values corresponding to similar behavior in one snapshot. So a cluster-based method has been paid more attention to extract the uniform region from raster datasets (Wemmertet al., 2009; Gunnermannet al., 2012). To better group discrete lattices into spatial clusters, the space similarity and spatial coherence are taken into account simultaneously, which have been reported by several documents (Makrogianniset al., 2005; Wanget al., 2009). In these approaches, spatial constraint,i.e., spatial neighborhood, is incorporated into cluster methods to avoid the noise and edge influence (Tolias and Panas, 1998; Liewet al., 2003). Generally, a cluster kernel in spatial pattern with a concise spatial extent and boundary is easy to identify and define, so a cluster-based method with a kernel expansion with spatial neighborhoods is more attractive to extract the sensitive space regions.

As for representation for the sensitive regions, as with both discrete object and continuous field properties, any purely object-oriented or field-based model has great challenges to represent them. The object-oriented model is to model real world objects with a precise and ‘crisp’spatial location and extent (Egenhoferet al., 1999), whereas, the field-based model is used for modeling physical properties,i.e., sea surface temperature, ocean color,etc., whose magnitude is dependent on its spatial location (Goodchild, 2009). In contrast to pure object- or field-like phenomena, most geographic phenomena have both object and field characteristics. This suggests the need for an integrated object-field representation (Yuan, 2001; Kjenstad, 2006; Goodchildet al., 2007). Many object-field models have been proposed to represent and analyze the complex geographic phenomena, such as a three-domain representation for spatiotemporal queries (Yuan, 1999), a geographical object with triple (Camaraet al., 2000), a hybrid concept of object fields (Cova and Goodchild, 2002), which all provide the foundation for designing the integration of field object model to represent and analyze the marine sensitive region.

The purpose of this paper is to propose a novel method for extracting marine sensitive regions with a kernelbased expansion algorithm and for representing them with a Field-object model. The paper is organized as follows. In Section 2, some basic concepts about marine sensitive factor, marine sensitive region and marine sensitive object are introduced. Section 3 presents a novel method for extracting marine sensitive regions and for representing marine sensitive objects. Section 4 gives a case of study to test the effectiveness and efficiency of the above method, while Section 5 presents our conclusions.

2 Three Basic Concepts

Marine sensitive factor (MSF)： MSR is a marine environmental parameter that sensitively responds to or drives global climate changes (Guo, 2009). It is also named essential climate variable by Global Climate Observing System, which may be one of sea surface temperature, sea level, ocean color and sea ice (Hollmannet al., 2013).

Marine sensitive region (MSR)： As one of the strongest signals to global climate changes, and among them involving various parameters acting as drivers and responders (McPhadenet al., 2006), ENSO is selected to identify the MSR. MSR is a region where a marine sensitive factor has both the similar characteristics in space and the similar behavior to that of ENSO in time series.

Marine sensitive object (MSO)： MSO is an object that represents the marine sensitive region with both a crisp spatial domain and changing values dependent on spatial locations. With the properties both a discrete object and a continuous field, a field-object model, the integration of object-oriented and field-based model, is used to represent MSO in this paper.

3 Methods for MSO Extraction and Representation

Fig.1 A workflow of MSO extraction and representation from long term remote sensing images.

EOF aims at decomposing the spatiotemporal co-variability into independent time components and spatial patterns. The dominant spatial pattern is to reveal the main spatial distribution with the similar variation in corre-sponding time components, so the uniform changing regions within spatial patterns are candidates for sensitive regions. After removal of seasonal patterns, an ENSO-related time component is identified from time components using correlation analysis with ENSO index, and then the corresponding ENSO-related spatial pattern is extracted. With the convergence of positive (negative) variations of spatial patterns, a kernel-based expansion with spatial neighborhoods is more effective and efficient for detecting MSR. The MSR has the precise spatial boundary, whereas the value within MSR varies dependently on its spatial location. Then an integrated Fieldobject model is used to represent their precise spatial extend and their continuous varying values simultaneously. Thus, the workflow of extracting and representing MSO from remote sensing images includes data pretreatment of long term datasets, ENSO-related spatial patterns extraction, MSR identification and extraction, and MSO representation as shown in Fig.1.

3.1 Marine Remote Sensing Images Pretreatment

Long term marine environmental parameters retrieved from remote sensing images have the main characteristics of seasonal variations, which are mainly dominated by solar radiance. Although these seasonal patterns are already well-known, under the background of global climate changes, the spatiotemporal patterns deviating from normal seasonal cycles get more attention on anomalous climate events analysis. How to remove the seasonal components of the climate time series before identifying the anomalous climate events in global change is a key issue.

Standard monthly averaged anomalous algorithm, which is also denoted as Z-Score (Zhanget al., 2005), takes all sets of values for a given month, from January to December, from long-term images, calculates the mean and standard deviation for that set of monthly values, and then standardizes each value by subtracting off the mean and dividing the standard deviation, as shown in formula (1). With few prior-knowledge, the Z-Score algorithm is more suitable for removing seasonal fluctuations.

whereidenotes a year,ja month from January to Decemberandδjare thejthmonth averaged value and standard deviation, respectively,andare the raw and transformed value (i.e., monthly anomalies) of the long term images, respectively.

3.2 ENSO-Related Main Mode Extraction

The interaction between marine sensitive factors and their relationships with ENSO constitute a complicated system. Among the independent spatial patterns and temporal components of marine sensitive factors derived from EOF, at least one of them is dominated by ENSO mechanism, such as sea surface temperature (Zhanget al., 2001), sea surface chlorophyll-aconcentration and marine primitive productivity (Wilson and Adamec, 2001), and sea surface level anomaly (Chenet al., 2010). The mode closely related with ENSO is defined as ENSO-related mode. On the basis of the EOF results, and in combination with ENSO index, the identification and extraction of ENSO-related mode are follows：

1) According to the significance testing for eigenvalues (Northet al., 1982), to determine the EOF modes, which have physical meanings.

2) For each time component with a physical explanation, to calculate the correlation coefficient between the time component and ENSO index.

3) If the correlation coefficient is above the significant confidence level, to obtain the main cycle of the time component through power spectrum analysis.

4) If the main cycle is about 4-7 years, the same as ENSO, to identify the time component as the ENSO-related one, and the corresponding spatial pattern as ENSO-related one.

3.3 Kernel-Based Expansion Cluster for MSR Detection

The regions covered by the positive maximum/negative minimum values of the spatial pattern are the positive/ negative changing center of marine environmental elements in 2-dimensional space, and these regions have the similar behavior to that of ENSO in phase/anti-phase. As the values around the changing center have a characteristic of aggregation, a positive/negative changing center can be defined as the cluster kernel. The cluster-based method with kernel spatial expansion comprises four key steps,i.e., cluster kernel definition, kernel-based spatial expansion, cluster mergence and MSR identification based on clusters.

3.3.1 Kernel definition

Generally, there at least exists one positive or negative changing center in the spatial patterns derived from the standardized EOF. The changing center reveals the main variation characteristics. The more the values closing to the changing center, the spatial characteristics are more apparent. To define the cluster kernel from the changing center needs a threshold value, which is a key issue. If the threshold value is too larger, the typical space range is smaller, and the potential changing center may be omitted. And if the threshold value is too small, the large amount space is regarded as changing center. Theoretically, the dynamic threshold value is more suitable for identifying cluster kernels. In this paper, the statistics of spatial principle pattern is used, which ensures at least one cluster kernel to exist,e.g., by sorting the values of spatial pattern of SST anomaly, the 3% the maximum positive value is 0.868, and the 3% the minimum negative value is -0.857.

3.3.2 Kernel-based spatial expansion

Kernel-based expansion mainly follows a spatial neighborhood. The 4-/8-neighborhood (Fig.2) is the generalspatial template in Geo-informatics for neighborhood analysis (Theobald, 2009). Fig.2a emphasizes on the horizontal and vertical characteristics and Fig.2b emphasizes on the diagonal one, when the space is homogeneous, Fig.2c is more suitable for spatial expansion. The cluster spatial neighborhood is defined byn-neighborhood of the cluster edge, and an example with 8-neighborhood is shown in Fig.2d.

Fig.2 Spatial neighborhood. (a) and (b) are 4-neighborhoods; (c) is 8-neighborhood; and (d) is a cluster neighborhoods.

According to the cluster kernel defined by statistics, a kernel-based spatial expansion algorithm consists of the following steps.

Step1： Initialize clusters from the predefined kernels and identify the clusters with an ID from zero.

Step2： Calculate the mean and standard deviation of each cluster.

Step3： Define the spatial neighborhoods of each cluster according to neighborhood types.

Step4： Calculate the averaged Manhattan Distance between a neighborhood and its cluster according to the following formula：

whereVNbhis a value of the neighborhood,ViClusteris a value ofithlattice within the cluster, andNis the total number of the cluster’s lattices.

Step5： Confirm whether the neighborhood belongs to its cluster or not according to formula (3)： if less than the distance threshold, the neighborhood is regarded as a part of this cluster; if not, the neighborhood is not a part of this cluster. The discriminate function is the following formula：

whereσClusteris the standard deviation of the cluster’s lattices, which is changing during its expansion; andλis a ratio scale.

Step6： For each neighborhood of the cluster, repeat Step4 to Step5 until all neighborhoods are identified belonging to the cluster or not.

Step7： For each cluster, repeat Step3 to Step6 until all cluster are expanded according to their spatial neighborhoods.

Step8： Update and replace the cluster with spatial expansion, and repeat Step2 to Step7, until all clusters have no expansions.

In the above workflow, how to define the distance threshold,λσClusteris a key issue. As the cluster is changing during the process of spatial expansion, the static threshold will omit significant information. According to the statistics of cluster, the dynamic threshold,i.e., standard deviation, is adopted in this paper. As for the ratio scaleλ, the smaller, the less neighborhoods are expanded into the cluster,i.e., some related lattices may be missed, while the larger, the more neighborhoods are expanded into the cluster, which may include some unrelated neighborhoods. Thus,λis determined by experiments

3.3.3 Cluster mergence

During the expansion process, when a neighborhood belongs to the adjacent clusters at the same time, these clusters will overlap with each other. The overlapped clusters mostly belong to the same change type,i.e., the positive or negative change. There is no sense that a positive changing region and a negative one merge in real world, as the transition region will make them independent of each other. As the same characteristics in variation, the overlapped clusters are merged into a new cluster. Its ID is assigned with the primary cluster with smaller ID, and the other cluster ID will be deleted. Fig.3 gives a merging example of the two clusters with an 8-neighborhood.

In Fig.3a, Cluster A and Cluster B are separate, even after the first spatial expansion, and they don’t join each other with premise of all neighborhoods belonging to their clusters. After the second spatial expansion, the new cluster A and B join with the same three neighborhoods (Fig.3b). With regard to the 8-neighborhood connectivity, the three common neighborhoods belong to the Cluster A and Cluster B at the same time, so the Cluster A and Cluster B will merge into a new Cluster C. Fig.3c is an example of the three common neighborhoods all belonging to the Cluster A and Cluster B.

After the changing center identification, kernel spatial expansion and cluster mergence, the spatial changing region can be obtained from the spatial patterns. On the basis of the clustering results, the MSR identification isan easy task. According to cluster ID, each cluster may be regarded as a MSR.

Fig.3 An example of adjacent cluster merging process.

3.4 Field-Object Triple for MSO Representation

As MSR representing a group of lattices with similar values in space and similar behavior in time, the objectoriented model is more suitable for describing its precise and ‘crisp’ spatial location and extent. However, the value within the MSR always varies depending on its spatial location. The MSR representation only with object-oriented model will lose large amount of detailed information.

As MSR with both object- and field-like properties, the integration of object-oriented and field-based model, a Field-object model, is proposed in this paper for organizing and representing them. The conceptualization framework of the Field-object model is in a triple form as ＜O, A, F ＞. Here O is an object in Euclidean space domain, which may be any type of vectors,i.e., point, line, polygon, vertex, multi-point, multi-line, multi-polygon,etc., and A is a set of attribute domains,i.e., marine sensitive factors in this paper. And function F assigns each location within the object O a value on the set of attribute domains A. In this framework, the discrete object is used to describe a precise and crisp spatial location and extent, while the continuous field is used to display how states of MSR vary in a space domain.

Fig.4 shows an example of the Field-object model for representing positive changing regions and negative changing regions. There exist four MSRs： one is a positive changing region, and the other three are negative ones. The object type,i.e., polygon, describes the precise spatial extend and its boundary of MSO, while the field model represents the internal structures of MSR heterogeneities. Simultaneous representation of the precise boundary and extend and the internal structures of MSO may be indicative of the physical dynamics in long time series analysis, which plays an important role in global climate changes and regional sea-air interactions.

Fig.4 An example of the Field-object model for representing marine sensitive region. (a) Field-object model for a positive changing region; (b) Field-object model for negative changing regions.

4 A Case of Study

The Pacific Ocean is located within 50°S to 50°N, and 100°E to 60°W, covering most of western Pacific Ocean warm pool and eastern Pacific cold tongue. It plays an important role in both global climate changes and regional sea-air interactions. In this region, marine dynamical, thermal and bio-optical parameters closely relate with each other. SST data from December 1981 to March2012 are taken as test data, which were obtained from the NOAA Optimum Interpolation Sea Surface Temperature V2.0 provided by the NOAA/OAR/ESRL Physical Sciences Division, Boulder, Colorado, USA, and available at http：//www.esrl.noaa.gov/psd/ (Reynoldset al., 2002). The Multivariate ENSO Index (MEI) based on the six main observed variables over the tropical Pacific and calculated as the first unrotated principal component (PC) of all six observed fields combined (Wolter and Timlin, 1993), was also used for correlation analysis and periodic analysis.

After the pretreatment of resample and Z-Score, the 364-month sea surface temperature anomalies, within a uniform framework with a spatial resolution of 1 degree and a time scale of a month, were used for EOF analysis. Correlation analysis between MEI and time components of EOF showed that the correlation coefficient between MEI and the first time component was up to 0.91 with a 99.9% confidence level, and they also had similar main cycles. The former was 3.5-6.0 year, and the latter was 2.5-6.5 year (Figure not shown), which were all within the ENSO periodical cycles with 2.0-7.0 year (McPhadenet al., 2006). So the first time component was identified as an ENSO-related time component, and the corresponding first spatial pattern was identified as an ENSO-related spatial pattern, as shown in Fig.5.

Fig.5 ENSO-related main mode from long term SST remote sensing image. (a) Time series of MEI and time component of first SST mode; (b) Spatial pattern of SST.

On the basis of the ENSO-related spatial mode, the dynamic threshold value according to the spatial statistics was calculated. On the assumption that the kernel region is greatly less than the final cluster,i.e., the kernel region might have little effects on final cluster, this paper selects 3% positive/negative value to determine the kernel regions. And the values within the kernel region have the similar and typical characteristics. When using 3%, the maximum positive values,i.e., the threshold value being 0.868, to define the positive variation centers, we got one kernel, and when using the 3% minimum negative values with the threshold value being -0.857, to define negative variation centers, we got 8 kernels, Fig.6(a) shows a total of 9 kernels. Then the total 9 kernels were initialized as clustering centers. After the recursive spatial expansion with a 8-neighbourhood and a discriminant ratio scale of 2.0, the Kernel 1 to Kernel 4 were merged into one cluster, Kernel 5 to Kernel 7 were merged into Cluster 2, Kernel 8 and Kernel 9 were expanded into Cluster 3 and Cluster 4, respectively, as shown in Fig.6(b).

Theoretically, variations of MSO have a positive/ negative phase relationship with ENSO indices. During the La Ni?a occurrence, SST anomalies within Cluster 4 are below multi-annual monthly anomaly, and SST anomalies within Cluster 1 to Cluster 3 are above multi-annual monthly anomaly, whereas the anti-events happen during El Ni?o occurrence. That is to say, SST anomalies within MSO may be a new indictor of or responder to ENSOevents.

In this paper, we extracted the multi-annual monthly SST anomalies of the time periods of the latest La Ni?a event occurring from August 2011 to March 2012, and used the Field-object model to represent the SST anomaly within MSO, and analyzed its spatiotemporal variations.

From June 2011 to July 2011, the onset of the latest La Ni?a event, the spatial averaged value of MSO was below its multi-annual monthly anomaly (Fig.7m), while the values in some locations within the MSO were not (Figs.7a and b). During the whole duration of La Ni?a, from August 2011 to March 2012, the spatial averaged values of MSO were all below their multi-annual monthly anomalies (Fig.7m), and the values of each location within the MSO were also below their multi-annual monthly anomaly, except for the shrinkage phase from February 2012 to March 2012, when some values within MSO were above the multi-annual monthly anomaly. After the La Ni?a event, the value within MSO began to increase above its multi-annual monthly anomaly (Figs.7k and l). Fig.7 also proves that MSO has similar properties to those of ENSO, and that MSO not only reveals its dynamic characteristics in time series, but also demonstrates its spatial variations in spatial domain.

Fig.6 Examples of marine sensitive object extractions from SST remote sensing images. Background is an ENSO-related spatial mode of SST. (a) Kernel defined with a threshold value of 3%; (b) Clusters of marine sensitive regions with an 8-neighbourhood spatial expansion and a ratio scale of 2.0.

Fig.7 MSO representation and its variations during the latest La Ni?a. (a) June 2011; (b) July 2011; (c) August 2011; (d) September 2011; (e) October 2011; (f) November 2011; (g) December 2011; (h) January 2012; (i) February 2012; (j) March 2012; (k) April 2012; (l) May 2012; (m) Time series of SST anomaly and MEI from June 2011 to May 2012. SST anomaly is a spatial averaged value within the MSO.

5 Conclusions

The advanced space-borne Earth observation technology makes it possible to conduct overall monitoring, analysis and simulation of the environmental parameters on the background of global change. On the premise of the known marine sensitive factors, this paper proposes a novel method for extracting marine sensitive regions using the long term remote sensing images and representing marine sensitive objects with a Field-object model. Using the sea surface temperature images during the period from December 1981 to March 2012, we also give a case of study to extract the marine sensitive objects about sea surface temperature anomaly, and analyze their spatiotemporal characteristics during the latest La Ni?a duration. Our main conclusions are summarized below.

1) On the basis of some basic concepts about marine sensitive factors, regions and objects, this paper designs a workflow from long term remote sensing images to extract marine sensitive regions with ‘data pretreatment→ENSO-related mode extraction→kernel identification→cluster expansion and mergence→marine sensitive region detection’. As one of the strongest signals of global climate changes, ENSO is absorbed into this workflow, and in combination with EOF result, the independent ENSO-related spatial distribution is derived. The statistics of spatial pattern is used to define cluster kernel, which ensures one kernel existing at least. The recursive spatial expansion and cluster mergence with spatial neighborhood integrate all the related lattices with similar variation in time series into one cluster, and make it different from others.

2) A Field-object model, the integration of ‘objectoriented and field-oriented’, with a triple form ＜O, A, F ＞is proposed to represent marine sensitive objects, where O means object domains, A means attribute domains and F means function domains. MSO with the triple form describes not only its spatial location and extent, also the spatial variations, which is more suitable for analyzing its dynamic characteristics in time series.

3) The marine sensitive objects of sea surface temperature are extracted, represented and analyzed using long term remote sensing images as a case of study. The dynamic analysis of one of marine sensitive objects during the latest La Ni?a duration shows that MSO has similar properties to those of ENSO. Both the spatial location and extent, and spatial variations representation also prove the effectiveness and the efficiency of the proposed method.

With the similar behavior to that of ENSO in time series, MSO may be a new indicator of or responder to ENSO events, which will have a scientific significance in both global change and regional sea-air interaction. So further research following this paper is underway to construct the long term datasets about MSO and develop framework for data mining and knowledge discovery to explore marine spatiotemporal association patterns under global changes.

Acknowledgements

This research was supported by the director projects of Centre for Earth Observation and Digital Earth (CEODE) (Nos.Y2ZZ06101B and Y2ZZ18101B), and the State Key Laboratory of Resources and Environmental Information System project, the National Natural Science Foundation of China (project No. 41371385), and the National High Technology Research and Development Program of China (project No. 2012AA12A403-5).

Camara, G., Montaero, A. M. V., Paiva, J. A., Gomes, J., and Velho, L., 2000. Towards a unified framework for geographical data models.Journal of the Brazilian Computer Society, 7 (1)： 17-25.

Chen, G., Wang, Z., Qian, C. C., Lv, C., and Han, Y., 2010. Seasonal-to-decadal modes of global sea level variability derived from merged altimeter data.Remote Sensing of Environment, 114 (11)： 2524-2535.

Cova, T. J., and Goodchild, M. F., 2002. Extending geographical representation to include fields of spatial objects.International Journal of Geographical Information Science, 16 (6)：509-532.

Egenhofer, M. J., Glasgow, J., Gunther, O., Herring, J. R., and Peuquet, D. J., 1999. Progress in computational methods for representing geographical objects.International Journal of Geographical Information Science, 13： 775-796.

Esaias, W. E., Iverson, R. L., and Turple, K., 2000. Ocean province classification using ocean colour data： Observing biological signatures of variations in physical dynamics.Global Change Biology, 6： 39-55.

GCOS, 2011.Systematic Observation Requirements for Satellite-Based Data Products for Climate(2011 update). GCOS Report 154, World Meteorological Organization, 43-58.

Goodchild, M. F., 2009. Field-based spatial modeling. In：Encyclopedia of Database Systems.Liu, L., and Ozsu, M. T., eds., Springer US, 1132-1138.

Goodchild, M. F., Yuan, M., and Cova, T. J., 2007. Towards a general theory of geographic representation in GIS.International Journal of Geographical Information Science, 21 (3)：239-260.

Gunnemann, S., Kremer, H., Laufk?tter, C., and Seidl, T., 2012. Tracing evolving subspace clusters in temporal climate data.Data Mining and Knowledge Discovery Journal, 24 (2)： 387-410.

Guo, H. D., 2009. Space-based observation for sensitive factors of global change.Bulletin of the Chinese Academy of Sciences, 23 (4)： 226-229.

Hannachi, A., Jolliffe, I. T., and Stephenson, D. B., 2007. Empirical orthogonal functions and related techniques in atmospheric science： A review.International journal of Climatology, 27： 1119-1152.

Hollmann, R., Merchant, C. R., Saunders, R., Downy, C., Buchwitz, M., Cazenave, A., Chuvieco, E., Defourny, P., de Leeuw, G., Forsberg, R., Holzer-Popp, T., Paul, F., Sandven, S., Sathyendranath, S., van Roozendael, M., and Wagner, W., 2013. The ESA climate change initiative： Satellite data records for essential climate variables.Bulletin of the American Meteorological Society, DOI： 10.1175/BAMS-D-11-00254.1.

Kjenstad, K., 2006. On the integration of object-based modelsand field-based models in GIS.International Journal of Geographical Information Science, 20 (5)： 491-509.

Liew, A. W. C., Leung, S. H., and Lau, W. H., 2003. Segmentation of color lip images by spatial fuzzy clustering.IEEE Transaction on Fuzzy Systems, 11： 542-549.

Makrogiannis, S., Economou, G., and Fotopoulos, S., 2005. A region dissimilarity relation that combines feature-space and spatial information for color image segmentation.IEEE Transaction on Systems, Man, and Cybernetics, Part B: Cybernetics, 35 (1)： 44-53.

McPhaden, M. J., Zebiak, S. E., and Glantz, M. H., 2006. ENSO as an integrating concept in earth science.Science, 314 (5806)： 1740-1745.

Milne, G. A., Gehrels, W. R., Hughes, C. W., and Tamisiea, M. E., 2009. Identifying the causes of sea-level change.Nature Geoscience, 2： 471-478.

Moron, V., 2003. Long-term variability of the Mediterranean Sea surface temperature (1856-2000).Comptes RendusGeoscience, 335： 721-727.

Nardelli, B. B., Colella, S., Santoleri, R., Guarracino, M., and Kholod, A., 2010. A re-analysis of Black Sea surface temperature.Journal of Marine Systems, 79 (1-2)： 50-64.

North, G. R., Bell, T. L., Cahalan, R. F., and Moeng, F. J., 1982. Sampling errors in the estimation of empirical orthogonal functions.Monthly Weather Review, 110 (7)： 699-706.

Oliver, M. J., and Irwin, A. J., 2008. Objective global ocean biogeographic provinces.Geophysical Research Letters, 35 (15), L15601, DOI： 10.1029/2008GL034238.

Reynolds, R. W., Rayner, N. A., Smith, T. M., Stokes, D. C., and Wang, W., 2002. An improvedin situand satellite SST analysis for climate.Jouranl of Climate, 15： 1609-1625.

Shen, S. H., Leptoukh, G. G., Acker, J. G., Yu, Z. J., and Kempler, S. J., 2008. Seasonal variations of chlorophyllaconcentration in the Northern South China Sea.IEEE Geoscience and Remote Sensing Letters, 5 (2)： 315-319.

Theobald, D. M., 2009.GIS Concept and Arcgis Methods. 4th edition. Conservation Planning Technologies, Colorado, USA, 346-347.

Tolias, Y. A., and Panas, S. M., 1998. Image segmentation by a fuzzy clustering algorithm using adaptive spatially constrained membership functions.IEEE Transaction on Systems, Man, and Cybernetics, Part A: Systems and Humans, 28 (3)：359-369.

Wang, Z. M., Soh, Y. C., Song, Q., and Sim, K., 2009. Adaptive spatial information-theoretic clustering for image segmentation.Pattern Recognition, 42： 2029-2044.

Wemmert, C., Puissant, A., Forestier, G., and Gancarski, P., 2009. Multi resolution remote sensing image clustering.IEEE Geoscience and Remote Sensing Letters, 6 (3)： 533-537.

Wilson, C., and Adamec, D., 2001. Correlations between surface chlorophyll and sea surface height in the tropical Pacific during the 1997-1999 El Ni?o-Southern Oscillation event.Journal of Geophysical Research, 106 (C12)： 31175-31188.

Wolter, K., and Timlin, M. S., 1993. Monitoring ENSO in COADS with a seasonally adjusted principal component index.Proceedings of the 17thClimate Diagnostics Workshop. NOAA/NMC/CAC, NSSL, Oklahoma Climate Survey, CIMMS and the School of Meteor., University of Oklahoma, Norman, OK, 52-57.

Yuan, M., 1999. Representing geographic information to enhance GIS support for complex spatiotemporal queries.Transactions in GIS, 3： 137-160.

Yuan, M., 2001. Representing complex geographic phenomena with both object- and field-like properties.Cartography and Geographic Information Science, 28： 83-96.

Zhan, H., 2008. Scaling in global ocean chlorophyll fluctuations.Geophysical Research Letters, 35, L01606, DOI： 10.1029/ 2007GL032078.

Zhang, Q. L., Weng, X. C., and Yan, T. Z., 2001. Spatial and temporal variations of SST field in the western Pacific warm ocean area.Oceanologia et Limnologia Sinica, 32 (4)： 349-354.

Zhang. P. S., Tan, P. N., Steinbach, M., Kumar, V., Shekhar, S., Klooster, S., Potter, C., 2005. Discovery of patterns in the earth science data using data mining. In：Next Generation of Data Mining Applications. Zurada, J., and Kantardzic, M., eds., IEEE Press, 167-188.

(Edited by Xie Jun)

(Received August 9, 2013; revised April 30, 2014; accepted June 3, 2015)

? Ocean University of China, Science Press and Spring-Verlag Berlin Heidelberg 2015

* Corresponding author. Tel： 0086-10-82178126 E-mail： xuecj@radi.ac.cn