Research On Object-oriented Optical Remote Sensing Image Feature Extraction And Classification

With the development rapidly of remote sensing earth observation technology,the spatial resolution of remote sensing data provided for surface coverage classification research is higher,the spectral information is richer,and the temporal resources are more complete.The traditional pixel-based classification technology cannot meet the needs of high-score image processing.The emergence of object-oriented analysis has played a positive role in solving the problems of "pepper and salt noise","same spectrum foreign body" and "same spectrum different spectrum" in image classification.How to use the rich feature information of image and further improve the classification accuracy is currently a hot issue in processing high-resolution images.In this paper,combining the above problems,the research of object-oriented optical remote sensing image feature extraction and classification is carried out.First,object-oriented segmentation is performed on the pre-processed high-score images to obtain image objects in multiple connected areas with high uniformity.Then based on the segmented image object,the feature library is constructed,the feature library is extracted,and finally the classifier integration is used to classify the image object,iterative research and analysis is performed according to its classification accuracy.1.In order to solve the problem of loss of geometric features of image objects caused by over-segmentation,the watershed segmentation algorithm of gradient reconstruction and iterative minimum calibration is proposed.First,using the Sobel operator for gradient calculation,considering the problem of multi-band remote sensing images and the different information between bands,the entropy values of each band are introduced,the ratio of the entropy value of each band in the full band is used as the gradient weight,then reconstruct the gradient;second,the improvement of the regional minimum threshold calibration,the foreground and background calibration,iterative calculation of the average of the foreground and the background,to obtain the minimum threshold,complete image segmentation.After comparative analysis of the results,the method provided in this paper is ideal for segmentation.It can solve the problem of over-segmentation and ensure the edge information of the features as much as possible,which can better support the subsequent analysis and detection of features.2.Aiming at the problem that the current integrated classification technology is mainly based on homomorphic integrated classification,which can not effectively combine different types of classifiers,the cascade generalization model based on weighted reconstruction of base classifiers is proposed.First,the training set is equally divided,and the training set and the test set are combined using the cross-validation method to train the multiple classification algorithms constructed,multiple classifier models are generated,at the same time,the weights are confirmed according to the accuracy of the cross-validation.Remove the value with too large difference in accuracy,and then take the average.After all the base classifiers obtain the correct average,the ratio of the sum of the average of each base classifier and the overall mean is the weight of the base classifier,the output The classification results and corresponding weights are used as new training data to train the classification algorithm to generate a metaclassifier,and then use the entire training set to train multiple classifiers to obtain a base classifier,and then form a base classifier weight Structured cascading generalization model.Finally,the reconstructed cascading generalization model is used to classify the image objects.By comparing and analyzing the classification results of the related classification algorithm and the cascading generalization model before reconstruction,it is found that the reconstructed cascading generalization model makes greater use of the advantages of each classification algorithm,whether it is from overall accuracy(OA)or Kappa coefficient,To a certain extent,improve the classification accuracy.The object coverage classification problem of remote sensing images is a relatively large and cumbersome task,and the object-oriented classification idea has opened up a new classification path for it.Image segmentation,feature extraction,and classifier selection have always been necessary steps for object-oriented remote sensing image classification.This paper is in possession of reference significance.