Application Of Radial Basic Function Neural Network On Land Classification

In recent years, With the quick development of 3S technology, the capturing, analysis andapplication of spatial information have come into a brand-new age, as an important means of spatialinformation capturing, remote sensing technology offer a large number of macroscopic, synchronicaland first-hand information which serves as land resource survey, land management in the short time)With the speedy development of obtaining remote sensing information technology, and it becomes amore and more strong capacity of offering data. it is very important for resource survey and landmanagement to extract the remote sensing technology information or remote sensing data)The paper took the bei mu town in the neijiang city as the region of trial(land type includingpaddy field, dry land, body of water, construction land, road land), chose ETM of LandSAT's imageas basic data, on the basis of analysis of remote sensing software ERDAS IMAGINE.data handlingsoftware MatLab, combining open country survey, designed the method of compored to classificationof remote sensing result of classification and ISODATA of traditional, MLC which base on radialfunction neural network, according to classification of image, calculated to classification matrix ofconfuse, accuracy assessment parameter, classification accuracy assessment parameter includingaccuracy of user, accuracy of produce, coefficient of Kappa, and according to classification accuracyassessment parameter carry out accuracy analysis, accuracy assessment, discussed the reason for faultclassification, miss of classification.The design of classification algorithm in the research, it is mainly divided into 5 party contentsincluding the choice region of training, the establishment of RBFNN algorithm, The simulation ofRBFNN algorithm, the drawing of classification image, the accurate assessment of classification result.Among them, the choice of trial region mainly adopted the method of open country and prior toestimate indoor, selection of typical land path build on the training region.Consturction of RBFNN,input the value of remote sensing in district of train to RBF network, and training, marked thettrainednetwork as net. the number of RBF layers' notes 6, equal to the number of remote sensing band, thenumber of output layers' notes 5, equal to the number of land type. Simulate of RBFNN, used functionsim, input value of remote sensing to RBF network, and output result, marked as output. The paperdrew classification image, made use of text-remote image switch tool, changed result of output toclassification remote image, accurate assessment and analysis of Classification, calculation to classification matrix of confuse, application to accuracy of user, accuracy of produce, coefficient ofKappa index, carry out accuracy analysis and assessment. Experimental results show that:On the basis of ERDAS IMAGINE, MatLab software, this paper built a type of technology ofcomputer auto classification of remote image which includes open countrysurvey, building on trainingof region, the establishment and simulate of RBF network, drawing of classification image, accuracyanalysis etc procedure.Remote land classification accuracy of RBFNN algorithm compared to ISODATA, MLC, thetotal Kappa coefficient respectively raise 0.1708, 0.0607, and total classification accuracy is equal to0.7933, quality of classification is fine, classification accuracy range from 0.60 to 0.80, closed to thebest(classification accuracy range from 0.80 to 1.00).Base on RBFNN remote land classification, classification area accuracy of land use type rangefrom 0.73 to 0.82, both dry land and paddy field is 0.78, construction land is 0.71, road land is 0.87,body of water is 0.83. Error rate of overall area should be the sum of the weighted average which isthe error absolute value in all land type, the value is 22%, the accuracy of overall area is 0.78.Base on RBFNN remote land classification, classification accuracy of land use type range from0.73 to 0.82, in accuracy of user, road land is 0.79, body of water is 0.73, construction land is 0.76, dryland is 0.82, paddy field is 0.80. In accuracy of produce, road land is 0.74, body of water is 0.73,construction land is 0.81, dry land is 0.81, paddy field is 0.81, quality of classification is fine,classification accuracy range from 0.60 to 0.80, very fine range from 0.80 to 1.00, result show thatRBFNN extracts information of land use type very well.