Remote Sensing Classification Of Forest Vegetation Based On Artificial Neural Network

Remote sensing classification is one of indispensable parts of forest resource survey and supervision. The classification accuracy directly influences the applicable level and practical value. It is a key issue of remote sensing image research to resolve the problem of multi-type images identification and meet the accuracy demand, and it has important significance. Based on the analysis of international remote sensing classification research, using Landsat7ETM+ data and geographical assistance data, applying BP neural network, the image of Manhanshan forestry center as a case study is classified. Its classification accuracy is compared with traditional method by statistical pattern (unsupervised and maxim likelihood method). The results are as follows:(1) Taking the multi-spectrum of LandSat7 ETM+ remote sensing image as main data source, and the present distribution map of forest resource of study area as geographical assistance data, remote sensing image of forest in study area is classified using BP neural network.. The total type accuracy of classification result is 67%, the total quantity accuracy is 84.65%, and the KAPPA coefficient is 0.6455. It show that the classification quality is better.(2) The type and quantity accuracy are compared and analyzed between BP neural network, maxim likelihood, simple and complex unsupervised classification method. The result shows that the total type accuracy of BP neural network respectively increases 7% 32% and 33% compared with the previous three methods, and total quantity accuracy increases5.3%. So, this classification method combined with BP neural network and geographical assistance data is an effective approach.(2) Based on the whole analysis of the subject map in study area, field survey data, text data, integrated with identification of experts, a method with a criterion file combined with RS and GIS as reference data is advanced. The result indicates that this method is a good classification way with high accuracy and efficiency.(3 ) Neural network method is more convenient to join the assistance data to classify than maxim likelihood method. The classification accuracy can be improved by making the best of the geographical assistance data provided by GIS.(4) Through repeated excise, remote sensing data and geographical assistance data (present distribution map of forest resource) are input, and three layers of BP neural network model of forest vegetation type is output. When the number of hidden network node is 16, the learning rate77=0.1, the momentum factor a =0.8 and exercise time is 8000, the classification accuracyof network model reaches the accuracy of 96%. It can meet the demand of remote sensing classification of forest vegetation using BP neural network.