Study Of Remote Sensing Image Classification Based On BP Neural Network

Using remote sensing method for geographic information has become one of themain content of geographic information, the most important contribution of thecontinuous developing of remote sensing technology is providing us with multi-date,multiband, multi-scale mass data, how to make these mass data play a role of theadvantages in our production and the life is what we need to think about. In the studyof remote sensing image classification, the traditional classification method such asminimum distance method, maximum likelihood method, spectral Angle classificationthough overcome the visual interpretation such as long time, easily influenced bypeople’s own deficiencies, but they don’t have good flexibility, high precision ofvisual interpretation and fail to give full play to the knowledge of expert. Therefore,In combination with visual interpretation methods and professional knowledge ofexpert experience and constitute a classification expert system to solve the ‘withdifferent spectrum with foreign body’ phenomenon in order to improve theclassification accuracy has become one of the hot spots.Artificial neural network classification method is the most widely used combinedwith visual interpretation methods and professional knowledge of expert, throughpractical experience, Error Back Propagation algorithm, hereinafter referred to as theBP neural network is one of the most applications in remote sensing imageclassification. BP neural network algorithm has good flexibility, the ability ofcomprehensive analysis, it can well-fitting nonlinear data that exist in the remotesensing image, so it can better solve ‘with different spectrum with foreign body’phenomenon, quickly and accurately identify and extract the feature information fromthe remote sensing images, so this is one of the traditional remote sensing imageclassification methods can’t achieve. In this paper, select Landsat-7remote sensing data and then pretreatment,including radiation correction, geometric correction, the optimal band combinationimage cropping, etc., the normalized processing of remote sensing images and DEMand then to. At the same time, build a BP neural network consists of three layers,determine neural network related parameters, to form a good classification model.Then determine the classification system and standard through field investigation andthe related research of geographic information data, select training sample region tothe training of BP neural network, when it is necessary to modify the parameters,input remote sensing and DEM data input in the ENVI software BP neural networkclassifier and then output the classification results and accuracy assessment. Finally,compare with traditional classification method to study the advantages anddisadvantages of BP neural network.Research shows that the BP neural network classification is better than thattraditional classification method, verify that the BP neural network for remote sensingimage classification is an effective classification method, illustrates that the DEM andother geographic information data to optimize the BP neural network has a significantrole. On the other hand, because of the limitation of time and experimental conditions,there are many aspects need further research, such as BP neural network is affected bythe number of samples, the prediction ability and the ability to learn there are conflicts,unable to clear number of hidden layer nodes, shows that the BP neural network havea lot of room for improvement.