Classification Based On Multi-source Data Wetlands Space Evidential Reasoning Knowledge Discovery

Wetlands are important natural resources and ecosystems, as they can alter floodwaves and provide essential habitat to many wildlife species. Wetlands are also vitalto global carbon and methane cycles. To protect and manage wetlands degradation, itis important and urgent to develop effective approaches for wetlands monitoring andmapping at large scale. With the development of remote sensing technology, theaccessible multisource data are increasing and this provides new opportunity andchallenge for geospatial information extraction of land water crossed wetlandsecosystem. In this paper, an evidential reasoning (ER) approach in data mining field isintroduced for wetlands monitoring and mapping. The ER approach can break throughthe limitations of traditional probability theory based classifiers and can integrate allfour types of geospatial data to discriminate wetlands accurately. Due to the generalityof ER approach, some basic scientific problems need to be resovled before applicationof it for supervised classification. This paper performed some related basic worksinnovatively, and some new algorithms generating evidence measure fromquantitative (including ratio and direct data) and qualitative data (nominal data), andan index for evaluating the utility of multisource data to supervised evidentialclassification were presented. These algorithms and the index were validated inseveral representative wetland areas. A multi-temporal remote sensing image basedwetlands mapping and knowledge discovery technique system was established byusing ER approach, and the classification result was compared with others methods.Mainly contents and conclusions of this paper were listed as follows:1. Evidence measure (EM) is the basic of ER theory. Therefore, developingreasonable and effective methods to derive evidence measure from original data is acrucial step in supervised evidential classification. Herein, a minimum distance andmodified frequency distribution methods were proposed to calculate EM fromquantitative and qualitative data and the methods were compared to traditionalalgorithms for supervised evidential classification. The results have proven the utility of our algorithms.2. Well know accurate classification of wetlands (land water crossed complexarea) is difficult. Multi-temporal satellite image data provide exclusive andcomplementary information on wetland discrimination, and thus have the potential forimproved wetland mapping. This paper introduced an evidential reasoning approachto combine multi-temporal data for wetland mapping, and established a techniquesystem. Compared with other classifiers, the ER approach was found to be muchbetter for wetland mapping based on multi-temporal data. Additionally, the knowledgeprovided by ER approach also holds the potential for improved wetland mapping andthis is an advantage of ER to others.3. Terrain data is a key geographical factor controlling the distribution ofwetlands, introducing terrain data to improve wetland mapping is, thus, necessary andinvaluable. Herein, a method by combinating linear function and knowledge wasproposed to derive evidence measure from terrain data for supervised evidentialclassification. The classification result was compared with traditional methods (MLC).Results showed that terrain data can be incorporated into evidential classification andimprove classification accuracy effectively. However, traditional MLC method cannot handle terrain data since unnormal distribution of trrain data. The classificationaccuracy of MLC decreased when terrain data added. Additionally, it was found thataltitude data contributes larger than slope data to evidential classification in this work.4. The data source used in evidential reasoning theory should be independent,therefore how to evaluate the contribution of multisource data to supervised evidentialclassification quantitatively is a key issue of the theory. In this paper, we proposed anew index to evaluate the contribution of multisource to supervised evidentialclassification based on the variation of evidence measures. The index was validatedby examing the relationship between classification accuracy and the index. Resultssuggested that the index can reflect the effectiveness of evidence accumulation, andthe variation feature of the index was consistent with the accuracy of classification.But the index can only reflect the variation of evidence measure, but not the accuracyof classified class. In general, this paper introduced an evidential reasoning approach to combineremote sensing image and geographical ancillary data for wetland mapping andknowledge discovery, and the method has improved wetlands mapping. The proposedevidence measures generation algorithms for quantitative and qualitative data haveenriched geospatial information extraction methods and are expected to accelerate thedevelopment of the field.