Data System

INTEGRATING MULTI-SOURCE INFORMATION VIA FUZZY CLASSIFICATION

METHOD FOR WETLAND GRASS MAPPING

X. Zhao a b, X. Chen a c *, A. Stein b

a

State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University,

Luo Yu road 129, China - ***@*****.***.***.**

b

International Institute for Geo-Information Science and Earth Observation, Hengelosestraat 99, Enschede,

Netherlands (xzhao,stein)@itc.nl

c

The Key Lab of Poyang Lake Ecological Environment and Resource Development, Jiangxi Normal University,

Nanchang, Jiangxi, China

Ths-17 Geo-information contribution to sustainability indicators

KEY WORDS: Indicator System, Geo-information, Remote Sensing, Monitoring, Synthesis Analysis

ABSTRACT:

Late greening vegetation has been regarded as significant indicator of flood recessional wetlands ecosystem in the Poyang lake

natural reserve (PLNNR). Mapping the wetlands, especially the distribution of late greening grassland is of great importance for

PLNNR managers and decision makers either for ecosystem dynamic monitoring or habitat sustainability assessment. The aim of

this paper is to explore the use of fuzzy classification methods to map wetlands land cover and to better represent the vegetation

landscape. The proposed fuzzy rule-based system integrates information on NDVI, wetness and elevation and expert knowledge

about vegetation growth condition and phenology. Nine types of land covers were classified, with four belonging to wetland

vegetation. A traditional error matrix analysis has been used for accuracy assessment. The results show that fuzzy classification

provides more detailed information on the possibility of vegetation presence, rather than presence/absence information obtained

from a crisp classification. Fuzzy classification thus best represents vague objects and is less sensitive to poor data quality.

classify early and late greening grassland(Tan 2002; Si 2006;

1. INTRODUCTION

Zeng 2006). Due to vague boundaries and highly heterogeneous

The seasonal dynamic pattern of wetlands ecosystem within of wetland grasslands, conventional crisp classification methods

Poyang Lake National Nature Reserve (PLNNR) is influenced lack capability of providing satisfactory accuracy. Fuzzy

by climate (rain fall, temperature and solar radiation), classifiers based on the concept of the fuzzy theory (Zadeh

environmental conditions (soil moisture, terrain), and 1965), are able to better deal with imprecise, uncertain or

hydrological fluctuation of the five up-rivers and the Yangtze ambiguous data sets and knowledge, than crisp classifiers.

River. Every year, various wetland vegetation areas green up in

spring. Some remain green throughout the summer flooding Fuzzy logic and fuzzy set theory have been applied in a variety

months (June to August), but become senescent in autumn of ecological mapping and prediction applications. Application

(September to November). We call these early greening examples include: fuzzy modelling of vegetation dynamics

grassland. Other wetland vegetation submerges during the (Foody 1996; Hajj et al. 2007), eutrophication in a lake (Chen

flooding period in dormancy, but starts to germinate at the end and Mynett 2003) and soil landscapes (McBratney and Odeh

of September when flood water ebbs and soil is exposed 1997; Bruin and Stein 1998; Metternicht 2003).

gradually. We call these late greening grassland. Late greening

grassland provides habitats that are popular among migrating The aim of this paper is to explore the use of fuzzy

birds in autumn. The growth of late greening grassland is classification to map wetlands land cover and to best represent

sensitive to the recession onset, which later influences forage by the vague vegetation landscape in the PLNNR. More

the migrating birds. It has thus been regarded as a significant specifically, we use fuzzy inference to recognize vague patterns

indicator of this flood recessional wetlands ecosystem(Liu and in a set of spatial data derived from Landsat TM image and an

Xu 1994). Therefore, mapping the wetlands, especially late elevation model. We argue that fuzzy inference, in combination

greening grassland distribution is of great importance for with multi-source information, preserves the vague nature of the

PLNNR managers and decision makers either for ecosystem physical wetlands vegetation landscape in the final mapping

dynamic monitoring or habitat sustainability assessment. result.

The grass wetland within PLNNR is dynamic, but its extent and 2. METHOD

presence are vague in space. Different vegetation communities

Fuzzy classification in spatial applications uses multi-source

are distributed in discernible ring belts around lake areas from

information. For example, it may include expert knowledge,

low to higher elevations (Tan 2002; Wu and Ji 2002). Previous

DEMs and remotely sensed image. We apply Mamdani's fuzzy

studies on this area combined multi-source information, to best

* Corresponding author. This is useful to know for communication with the appropriate person in cases with more than one author.

1463

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008

inference system to map the wetlands land covers in the 2.2 Fuzzy sets, membership function and inference system

PLNNR. A brief review of the definition of fuzzy set,

Fuzzy set theory has been applied in various fields, such as

membership function and fuzzy inference system is given first,

image clustering (Binaghi et al. 1999), decision making and

which is followed by a fuzzy rule-based system applied to the

control. Basic definitions of fuzzy sets and fuzzy logic can be

land cover classification in the study area.

found in(McBratney and Odeh 1997; Fisher 2000). A fuzzy set

A is characterised by a membership function (MF), A(x), that

2.1 Study Area and data sets

assigns to each element x a grade of membership ranging from

The study area - Poyang Lake National Natural Reserve zero to one.

(PLNNR) is approximately 240 km2 (29 05 - 29 18 N,

[0,1]

X

A = {x, A(x)} for each x A(x) (3)

115 53 -116 10 E). The area is located to the northwest

of the Poyang Lake in the JiangXi province, central China. It is

A(x) = 0 means that x does not belong to the subset A, A(x) =

situated within the middle range of the Yangtze River basin.

1 indicates that x fully belongs, and 0 65%}, and compare it with field samples, then we

Figure 5 shows the nine land cover distribution within the

observe a 74.55% overall accuracy and a kappa value of 0.48

PLNNR area, based on the fuzzy rule-based classification of the

Landsat TM image and DEM and expert knowledge on the

3.2 Uncertainty map and analysis

phenology of wetland vegetation. Pixels excluded by the 19

rules were assigned to unclassified. The membership values

A fuzzy classification has the advantage that a measure of

from all the land covers were compared and the class with the

uncertainty factor C can be defined for each pixel as

highest membership value was assigned to the pixel label. The

accuracy assessment for all the 9 classified land covers shows a

C = 1 max m(x) (4)

71.69% overall accuracy and kappa value equal to 0.68 (table 1).

This wetlands land cover map provides more detail location

This measure indicates whether or not the classification has

information on different wetland vegetations than previous

yielded a clear response. The highest value of uncertainty factor

studies (Zheng 2001; Zhao et al. 2003; Si 2006; Leeuw et al.

is 0.5, which indicates the unclassified pixels. The lowest

2007). Dominant wetland vegetation including late greening

occurring value equals 0.11 instead of zero, since membership

vegetation, early greening vegetation and aquatic vegetation

value of each output has not successfully captured 1. The

were differentiated from terrain vegetation. Moreover,

reason is that centroid operator was used for defuzzification of

germinant late greening vegetations which still do not appear

the system output. It confines the range of vigorous late

blooming in October and wet bare lands which have suitable

greening vegetation value from 0.89 to 0.1 and the range of

wetness condition for late greening vegetation shooting up have

other land covers value from 0.75 to 0.25. Membership values

been detected.

outside this range are impossible to be reproduced by fuzzy

inference. The system output variables were divided into four or

From the classification result, we notice that 108 plots have

two categories. The number of divisions has an impact on the

been classified as early greening vegetation, among which only

possible range of output values. A fine divisions corresponds

55 were correct, leading to 50.93% users accuracy. Since early

with a large range of output membership values. Considering all

greening vegetation grows in a wide range of elevations and

uncertainty factors, we considered pixels with C 0.35 as having a high

misclassified to it. Obvious misclassification happened

potentiality of misclassification. The uncertainty map in figure

between vigorous late greening vegetation and early greening

7 thus provides us with a good opportunity to highlight the

vegetation. The suitable growth elevation of early greening

deficiencies in the rule-based system and can be referenced as

vegetation ranges wider than that for late greening vegetation.

priori sampling locations for validation.

Therefore, they have distinct phenology but still are highly

mixed at the moderate elevation level. In Miscanthus +

Cynodon + Carex vegetation community, for instance, taller

Miscanthus and Cynodon grows in the superstratum and Carex

is the undergrowth (Wu and Ji 2002). Some germinant late

greening vegetation was misclassified as early greening

vegetation. The reasons are that they have similar NDVI values

of image pixels which reflect both of them appearing less

vigorous in the fields, and that they possibly mix in their growth

in lower elevation near the shortly exposed river bank. Some

terrestrial vegetation was misclassified as early greening

vegetation mainly due to the similar spectral characteristics and

suitable elevation of crop fields and early greening vegetation.

River bank area and wet bare land near the river both have a

low NDVI and a low elevation. They only differ in wetness

degree. Sample plots of river bank and wet bare land collected

from Landsat TM image are imprecise by nature, thus

underestimating the accuracy of the assessments on these two

classes.

1466

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008

Fig. 5. Wetland land cover map classified by fuzzy rule-based system

Reference

CT NC PA UA

Class Class Class Class Class Class Class Class Class

Classified

1 2 3 4 5 6 7 8 9

57 *-**-*-*-*-*-*-*-** 57 62.64% 67.86%

Class 1

**-**-*-*-*-*-*-*-*-** 55 66.27% 67.90%

Class 2

**-**-**-*-*-*-*-** 0 108 55 73.33% 50.93%

Class 3

3 4 *-**-*-*-*-*-*-**-** 57.33% 86.00%

Class 4

1 2 1 *-**-*-**-*-*-**-** 56.63% 68.12%

Class 5

2 0 2 6 *-**-*-*-*-**-** 77.19% 72.13%

Class 6

0 1 0 *-**-*-**-*-*-**-** 71.43% 67.16%

Class 7

0 0 1 0 0 6 0 68 0 75 68 79.07% 90.67%

Class 8

0 0 1 8 8 0 1 0 90 108 90 100.00% 83.33%

Class 9

Reference

**-**-**-**-**-** 63-86-90-703-***

Total

Overall Classification Accuracy = 71.69%

Overall Kappa Statistics = 0.6809

Table 1. Wetlands land cover map accuracy report (RT= reference total, CT= classified total, NC= number correct, PA= producers

accuracy, UA= users accuracy)

1467

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008

Fig. 6. Membership values of possibility of vigorous late Fig. 7. Locations of pixels with high uncertainty (in black)

greening vegetation presence

1.20 300

Proportion of late greening

y = 125.31x + 98.216

1.00 250

R2 = 0.7771

0.80 200

vegetation

NDVI

0.60 150

0.40 100

0.20 50

0.00 0

0.00 0.20 0.40 0.60 0.80 1.00 0.00 0.20 0.40 0.60 0.80 1.00

Mem bership value Mem bership value

Fig. 8. Correlation between membership values and vegetation proportions (a), membership values and normalised NDVI values (b)

membership values can be regarded to corresponding vegetation

4. DISCUSSION

proportion in the fields. In fact, output membership functions

describe possibility instead of proportion. A total of 109 field

4.1 Validation of soft classification result

sample plots with late greening vegetation proportion higher

Soft classification approaches based on fuzzy set theory or than 50% has been used to extract membership values on soft

probability statistic theory are expected to give more result map in figure 6. Figure 8(a) shows that the correlation

information than only yes or no by hard classification. How to between membership values and vegetation proportion is not

evaluate the soft classification result, however, has been significant. In the absence of a directly corresponding test

proposed as a research and practical question(Gopal and measurement, validation of the soft result becomes difficult. In

Woodcock 1994; Foody 1996; Binaghi et al. 1999; Townsend figure 8 (b), membership values and normalised NDVI values

2000). When using historical images and field collected data, as appear highly positively correlated. It shows that NDVI values

in this study, validation will be confined by the existing data can serve as indicator of late greening vegetation presence.

and soft results can not be fully assessed. Even if there is an Possibility values in figure 6, however, contain more

opportunity to collect test samples in the fields for a particular comprehensive information about vegetation presence than the

soft classification validation, we may question what kind of NDVI values, since it not only considers vegetation index

information in the sample point will correspond to the soft (NDVI) but also the wetness index (LSWI) and elevation. In

result. In this study, taking the vigorous late greening summary, it is no problem to compare different possibility

vegetation output as an example, membership functions to map values and to select the land cover with maximum one as

possibility degree of vegetation presence have values between classification label, but to use the single soft result of one land

cover deserves more attention.

0 and 1. Membership values close to 1 indicate high possibility

to present, those close to 0 indicate high possibility of absence, 4.2 Limitations and further improvement

and those around 0.5 indicate highly uncertain decisions. The From the low users accuracy of early greening vegetation, we

uncertainty map in figure 6 thus gives a rough impression notice that fuzzy rule-based system works less efficiently if the

where high certainty happens, i.e. in the homogeneous classification object has loose constraints, as may apply for

vegetation zones, and where large uncertainty occurs, i.e. in widely distributed vegetation without growth limits,. Part of the

vegetation transition zones. This does not necessarily mean that

1468

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008

Boles, S., X. Xiao, et al., 2004. Land cover characterization of

classification error can be corrected by an image from another

Temperate East Asia using multi-temporal image data of

time. For example, in the late November image, early greening

VEGETATION sensor. Remote Sensing of Environment. 90, pp.

vegetation and germinant late greening vegetation can be

477-489.

classified with more certainty, when most of early greening

vegetation withered versus germinant late greening vegetations

blooming. The fuzzy rule-based system is flexible to add the Bruin, S. d. and A. Stein, 1998. Soil-landscape modelling using

multi-temporal image data, only at the price of additional fuzzy c-means clustering of attribute data derived from a digital

computing time. elevation model (DEM). Geoderma. 83, pp. 17-33.

The rules in the fuzzy rule-based system are based on expert Chen, Q. and A. E. Mynett, 2003. Integration of data mining

knowledge on the vegetation growth conditions and phenology. techniques and heuristic knowledge in fuzzy logic modelling of

Therefore, mapping natural land covers can benefit from the eutrophication in Taihu lake. Ecological Modelling. 162, pp.

extra information besides the image data. If existing knowledge 55-67.

about the classification targets is insufficient or implicated in

the data, the explicit rules definition used in this study can not Fisher, P., 2000. Sorites paradox and vague geographies. Fuzzy

applied any more. In addition, if the number of system inputs is Sets and Systems. 113, pp. 7-18.

large, there is no possibility to try out each possible inputs

combination. Other data mining methods, such as neural

Foody, G. M., 1996. Approaches for the production and

networks and optimizing methods such as simulating annealing

evaluation of fuzzy land cover classifications from remotely-

can be the alternatives for rules construction (Zheng 2001;

sensed data. International Journal of Remote Sensing. 17(7), pp.

Bardossy and Samaniego 2002).

1317-1340.

The fuzzy rule-based classification in this study is an

Foody, G. M., 1996. Fuzzy modelling of vegetation from

unsupervised classification. It integrates expert knowledge and

remotely sensed imagery. Ecological Modelling. 85, pp. 3-12.

multi-source data by means of fuzzy inference. Other

researches on precise membership function definition (Vlag and

Gopal, S. and C. Woodcock, 1994. Theory and methods for

Stein 2007) and rule importance measurement (Zheng 2001)

accuracy assessment of thematic maps using fuzzy sets.

can provide further improvements on this classification method.

Photogrammetric Engineering & Remote Sensing. 60(2), pp.

If more samples exist for supervised classification, the

181-188.

possibility values given to each land cover can be used as priori

probability instead of equal probability value in maximum

Hajj, M. E., A. Begue, et al., 2007. Multi-source Information

likelihood classification approach.

Fusion: Monitoring Sugarcane Harvest Using Multi-temporal

Images, Crop Growth Modelling, and Expert Knowledge. in

International Workshop on the Analysis of Multi-Temporal

5. CONCLUSION

Remote Sensing Images. Leuven, Belgium: IEEE.

Wetland ecosystem is characterised by high dimensionality,

complexity and non-linearity. Ecosystem knowledge is usually Hunt, E. R., 1994. Relationship between woody biomass and

semi-qualitative, and a large field dataset is difficult to obtain PAR conversion efficiency for estimating net primary

due to high sampling cost and inaccessible plots. This makes production from NDVI. International Journal of Remote

classification approaches solely depending upon expert Sensing. 15(8), pp. 1725-1730.

experiences rather subjective and let alone purely data-driven

approaches such as artificial neural network. The fuzzy rule- J. A.Gamon, C.B.Field, et al., 1995. Relationship between

based system applied in this study is a practical approach to NDVI, canopy structure, and photosynthesis in three

deal with semi-qualitative knowledge and semi-qualitative data. Califormian vegetation types. Ecological Applications. 5(1), pp.

This research shows that land cover mapping in the PLNNR 28-41.

wetland ecosystem can benefit from a fuzzy rule-based

classification method by combining expert knowledge, DEM,

Leeuw, J. d., Y. Si, et al., 2007. Mapping flood recessional

and remotely sensed image. This method provides more detail

grasslands used by overwintering geese: a multi-temporal

information on possibility rather than only presence or absence

remote sensing application. in 5th spatial data quality

information in a crisp classification. It has the powerful

international symposium Enschede, the Netherlands.

capability of representing vague objects and is less sensitive to

data quality problems.

Liu, Y. and Y. Xu, 1994. The study of influence and

countermeasure of Sanxia project to Poyang Lake's migratory

birds reserve Journal of JangXi Normal University. 18(4), pp.

REFERENCES

375-380.

Bardossy, A. and L. Samaniego, 2002. Fuzzy rule-based

McBratney, A. B. and I. O. A. Odeh, 1997. Application of

classification of remotely sensed imagery. IEEE Transaction on

fuzzy sets in soil science: fuzzy logic, fuzzy measures and

Geoscience and Remote Sensing 40(2), pp. 362-374.

fuzzy decisions. Geoderma. 77, pp. 85-113.

Binaghi, E., P. A. Brivio, et al., 1999. A fuzzy set-based

Metternicht, G. I., 2003. Categorical fuzziness: a comparision

accuracy assessment of soft classification. Pattern Recognition

between crisp and fuzzy class boundary modeling for mapping

Letters. 20, pp. 935-948.

salt-affected soils using Landsat TM data and a classification

based on anion ratios. Ecological Modelling. 168, pp. 371-389.

1469

The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008

Zeng, Y., 2006.Monitoring grassland in Poyang Natural

Si, Y., 2006.Mapping flood recessional grasslands grazed by

Reserve, China. MSc, Enschede, ITC, pp. 54.

overwintering geese: an application of multi-temporal remote

sensing. MSc, Enschede, ITC, pp. 60.

Zhao, X., M. Yuan, et al., 2003. A Study on Remote Sensing

Investigation and Comprehensive Utilization of Low-grassland

Tan, Q., 2002.Study on remote sensing change detection and its

in Poyang Lake Region. Journal of JangXi Normal University.

application to Poyang international importance wetland. PhD,

(1).

Beijing, Institute of Remote Sensing Application Chinese

Academy of Science, pp. 136.

Zheng, D., 2001.A Neuro-fuzzy approach to linguistic

knowledge acquisition and assessment in spatial decision

Townsend, P. A., 2000. A quantitative fuzzy approach to assess

making. Vechta, Germany, Vechta, pp. 156.

mapped vegetation classifications for ecological applications.

Remote Sensing of Environment. 72, pp. 253-267.

Vlag, D. E. v. d. and A. Stein, 2007. Incorporating uncertainty ACKNOWLEDGEMENTS

via hierarchical classification using fuzzy decision trees. IEEE

This work was funded by the 973 Program (Grant

Transaction on Geoscience and Remote Sensing. 45(1), pp.

No.2006CB701300), Sino-Germany Joint Project (Grant No.

237-245.

2006DFB91920), Open Fund of Shanghai Leading Academic

Discipline Project (T0102) and Key project of LIESMARS

Wu, Y. and W. Ji, 2002.Study on Jiangxi Poyang Lake National

(2006).

Nature Reserve.

Zadeh, L. A., 1965. Fuzzy sets. Information and Control. 8, pp.

338-353.

1470

Contact this candidate