Research On Multi-spectral Remote Sensing Image Classification Based On Formal Concept Analysis

With the rapid development of space technology, digital image processing and computertechnology, the remote sensing technology with a wide range of applications and prospects hasbeen further expansion and promotion. All kinds of math and Classification algorithms,whichaims to large numbers of remote sensing data bands, the strong correlation of bands, a hugeamount of computation, as well as the various uncertainties of the classification, have been moreand more applied to remote sensing image classification.As a mathematical tool of data analysis and implicit knowledge mining, Formal ConceptAnalysis theory have made a lot of results in many areas, such as machine learning, softwareengineering and knowledge acquisitioning. The mathematical tools, formal concept analysis hasbeen introduced in this thesis, established a formal context and analyzed features in the remotesensing image more rapidly and compactly, and effectively achieved image classification. Inorder to reduce the numbers of classification band, attribute containing theory of concept latticehas extracted the best band combination with a rich amount of information and lest correlation.There are two ways to analyze the classification features and extract the classification rules,constructing concept lattice and searching out partial order diagram. Finally considering of thedata processing features in formal concept analysis, the thesis formatted the classifier withobject-oriented multi-feature combination model, aiming to large dimensions of the classificationcharacteristics and the spots in classification results.This article focuses on the application of formal concept analysis in feature analysis andclassification rule mining of remote sensing image from point of view of the lattice theory, andaround this core issue, made some progress in algorithms and applications. Completed the majorwork are as follows,1. Considering of characteristics of data in formal concept analysis (discrete data), regardremote sensing image and its feature vector as a knowledge expression system, thebands and various features of the multi-spectral images as the attributes of the conceptand transform the multivalued background to a single valued background by using ofthe conceptional scale, then on this basis establish format background and data model.2. At the band selection stage, the design of the multi-spectral remote sensing image bandselection algorithm based on attribute reduction theory and experimental verificationhave been realized. The experimental results show that, there is more or less relativitybetween the bands of multi-spectral remote sensing images. More independent and informational bands can be extracted by the attribute (band) reduction of formal concept,to achieve reduction of the classification data.3. Studied the problem of feature integration in the sample training and introduced theclean (clear) conceptional background and GM partial order relations and so on.Considering of the relativity between the features and using of the methods of attributereduction and association rules extraction, construction of multi-feature model has beenrealized in phases of sample features extraction and analysis, and remove redundantfeatures to improve the efficiency of modeling.4. On the issue of extraction of classification rule, taking full advantage of the partial orderof concept lattice, designed an object-oriented algorithm，which aims to differentiatemany types of surfaces, and verified the algorithm by experiment.5. On the basis of the existing multi-feature model, achieved the classification of remotesensing image, compared with maximum likelihood classifier and given the accuracyanalysis. Analysis of experiment showed, the object-oriented feature model extracted inFCA is applicable to the supervised classification, and can achieve higher classificationaccuracy.