DOI: 10.36347/sjet

Pages: 482-486

Application of improved BP neural network (NN) model to research the ETM + remote sensing (RS) data classification. Select six of LANDSAT 7 ETM + multi-spectral bands as the data source, through read the six ETM image respectively in each pixel’s gray value, construct a six - dimensional input sample of the sample space: Vi={Vil,Vi2,Vi3,Vi4,Vi5,Vi6}. Through the experiment, ETM1, ETM3, ETM5, ETM7 was determined four wave bands as the input layer(the number of input layer neurons M = 4), and the geological object was divided into five categories(the number of output neurons N = 5), the optimal improved BP NN structure for the three-layers BP network, the number of the hidden layer nodes is 24, the input layer node number is 4, the output layer nodes is 5, excitation function for the S-type logarithmic function. The optimal network training parameters : the initial weights is 0.1, the learning rate is 0.9 , increasing rate of learning speed is 1.0 , decreasing rate of learning speed is 0.05 , the momentum factor is 0.001 , the network 's global error is 0.01 .Network model after 2000 times or less training, to achieve convergence. The model was applied to RS images of Mulei County of Xinjiang carry out geology type classification, the total classification precision reaches 89.06%, rapid classification quality is good, meet the production quality requirements.