Studies On Vegetation Classification Method Based On High-resolution Remote Sensing Image

Take state owned forest farm in Jiangle County of Fujian Province as the study area, two kinds of high resolution remote sensing images, ALOS and QuickBird images, were selected as the research object classification of remote sensing image feature. Firstly, this thesis introduces the theory of acquisition and remote sensing classification method of remote sensing data, and the preprocessing of remote sensing images; then the advantages and disadvantages of current commonly used remote sensing classification method-BP neural network method was analysed.The improvement method of the BP neural network method was put forward and analyzed. Selection method of remote sensing image texture factor was studied at last. As a new neuron, texture factor was added into the neural network classification method and experiented., which was analyzed and compared with previous classified results at detail. The main results are as follows:(1) The main function of the activation function is analyzed and the BP neural network algorithm is improved. In this paper, in order to select the condition of activation function, focuses on the analysis of the Cauchy distribution (Cauchy), the Laplasse distribution (Laplace) and the Gauss error function (Erf) as the activation functions and a new BP algorithm, using the XOR problem simulates the above 3 kinds of new activation function and the Sigmoid function (Logsig), the hyperbolic tangent function (Tansig) and normal distribution function (Normal) simulation test according to the results of the simulation analysis of the above 6 kinds of activation function and its influence on the performance of convergence in different parameters.Suitable incentive function was screened from the point of view of the range of the function itself and its derivatives and convergence angle, and then was used for classification of remote sensing image objects. Testing analysis of the conclusions can provide reference for feedforward neural network based on BP algorithm in the design of selective activation function.(2) After the improving neural network method, the LAP-BP algorithm was applied to the classification of ALOS images and QuickBird images. Four species were classified, mainly including Sugiki, pine, hardwoodand soft broad-leaved tree species. The other four land types were bare land, building land, road and water. Accuracy ALOS image classification results for the LAP-BP algorithm to classify the overall classification accuracy was 77.95%, Kappa coefficient was 0.7740,and the overall classification accuracy was the most ideal. Minimum distance classification overall accuracy was 73.90%, whose Kappa coefficient was 0.7343, and the accuracy was the lowest. Accuracy of QuickBird image classification results for the LAP-BP algorithm to classify the overall classification accuracy was 76.82%, whose Kappa coefficient was 0.7559 and, the overall classification accuracy was the most ideal. Maximum likelihood classification overall accuracy was 73.42%,whose Kappa coefficient was 0.7295,and the accuracy was relatively low.(3) This study put forward the quickly determining optimal or near optimal method of texture structure factor by using orthogonal test technology research. Through the combination of 44 texture factor, using 8 commonly used texture index to calculate of the texture image of high resolution remote sensing images.Separation index J value of screening the optimal combination that is suitable to describe isolation degree of different texture types according to the principle of class separability criterion. the optimum factor combination of texture:Windows 29*29, step 5, gray level 64, the direction of 45 degrees; QuickBird image texture, the separability of the largest J value, the best texture factor:47*47, step 3, gray level 128, the direction of 135 degrees.(4)Afrer selecting the best combination of factors from texture information, texture classification as the characteristic value was added to ALOS image and QuickBird image. Classified objects were still Sugiki, pine, hardwood and soft broad-leaved tree species and bare land, building land, road and water. Accuracy of ALOS image classification results for the LAP-BP algorithm to classify the overall accuracy was 83.59%.0.8305 is the coefficient of Kappa, and the overall classification accuracy is the most ideal. The minimum distance classification overall accuracy is 80.14%, whose Kappa coefficient is 0.7956, and it is the lowest classification accuracy. Accuracy of QuickBird image classification results for the LAP-BP algorithm to classify the overall accuracy was 85.36%. The coefficient of Kappa is 0.8494, and the overall classification accuracy is the most ideal. The overall accuracy using minimum distance classification is 79.81%,and Kappa coefficient is 0.7937, which is the lowest.