Research On Classification And Recognition Method Of Road Targets In Multispectral Remote Sensing Images Based On Machine Learning

With the application of remote sensing image more and more extensive the traditional use of artificial methods to detect and identify objects in the image has been far from meeting the needs of the present.Due to the large amount of remote sensing image data,complex target types,rich texture information,and large scene changes,the accuracy of current target detection methods is not very high,and often there will be missed detection,affecting the performance of target detection.Therefore,how to use the multi-spectral information of remote sensing images and adopt reasonable and effective processing methods to improve the detection and recognition performance of remote sensing images is a topic that researchers pay more and more attention to.Taking road target in multi-spectrum remote sensing image as the object,machine learning is used to study the effective detection and recognition of road target from three aspects,such as feature-based road target description,traditional machine learning road detection and modern machine learning road recognition.The main work of the thesis are as follows:(1)Road target description based on effective texture features.The Gray Scale Co-occurrence Matrix(GLCM)is used to effectively extract texture features to accurately describe road targets.By analyzing the contributions of different size samples,different band images and GLCM generation parameters to road targets,remote sensing images of RGB color space,G band and NIR band are firstly selected and GLCM is then determined.By comparing the discrimination of road and non-road targets based on the texture features calculated by GLCM,three texture features,Inverse Different Moment(IDM),Entropy(ENT)and Contrast(CON)are regarded as the most effective feature description of road targets for later detection and classification.(2)Road detection method based on improved SVM.A bilinear segmented dichotomy grid searching is proposed to optimize the parameter determination of SVM classifier based on RBF kernel.First,the penalty factors corresponding to the highest accuracy of cross-validation are found to determine the best search segment in the initial search range.Then,segmented dichotomy method is then adopted to solve the best parameters corresponding to the highest accuracy of each SVM segment iteratively.Finally,the optimal parameter corresponding to the maximum value of the highest accuracy of SVM is found,that is,the SVM model parameter optimization result.Compared with traditional methods such as bilinear method,grid search method and bilinear grid search method,the parameter optimization method proposed in this paper has the advantages of small training amount,simple calculation and high learning accuracy,so that improved SVM method has higher generalization performance.In remote sensing image road target detection,the accuracy rate reaches 70%.(3)Road target recognition based on Very Deep Convolutional Neural Network(VD-CNN).By adopting the superposition strategy of different small convolution kernel,the convolution layer with different depths,pooling layer and full connection layer are selected to reconstruct four CNN structures,including VD8,VD10,Lower VD10,and Highter VD10 based CNN network for road target recognition.Compare with the performances on training of the four networks and recognition accuracy,the reconstructed vd10-cnn network structure based on the superposition of small convolution kernel achieves the best results of road classification and recognition due to its reasonable depth,and the accuracy reached 93%.