Research On High-Resolution Satellite Image Classification And Change Detection Algorithms

With the rapid development of urbanization,the construction land of cities and town is increasing.At the same time,a series of rapid changes in the land surface such as occupation of cultivated land and forest land occur,and have attracted great attention from the land management departments at all levels.However,how to quickly,timely and accurately discover the changes and information of urban land use is critical,and the remote sensing change detection technology provides a scientific method to solve this issue.Through the high-resolution image,a rich data base is provided for accurate extraction of land surface change information.At the same time,the multi-resolution of the high-resolution image brings new difficulties and problems to the classical transform detection method.In the multi-resolution and multi-scale environment,same spectrum with different objects and different objects with the same spectrum,scale of the texture structure,reduction of interclass variance and the increase of intraclass variance,etc.of the high-resolution image are all difficult problems in high-resolution image change detection.Therefore,in this paper,the prominent problems in land surface change detection are focused on,and the research on the multi-scale change detection methods is conducted from aspects of ridgelet transform feature and fusion feature,convolutional neural network,etc.of the high-resolution image,aiming to comprehensively utilize the multi-resolution information of multiple high-resolution images,weaken the influence of errors in the preprocessing and detection process,so as to enhance the accuracy and rationality of the surface change detection results,construct a new change detection method,and provide the theoretical support for application of the change detection technology and production practice.The main research and innovation are as follows:Two high-resolution image classification and change detection processing technologies,including high-resolution image classification algorithm based on ridgelet and convolutional neural network and the change detection method based on spectral feature level fusion and multi-scale segmentation voting decision-making,are innovatively proposed.The classification algorithm on the basis of the ridgelet and convolutional neural network is based on the theory of ridgelet and convolutional neural network,which combines the "low level" simple features extracted through ridgelet and the "high level" features extracted through the neural network.In the process of feature extraction,the dependence of fusion features on the training set is reduced by using the "low level" features extracted through the ridgelet,making the fusion features more independent;In this process,the convolutional neural network suppresses the generation of noise,raises the consistency of classification regions,and finally improves the classification accuracy of images.In the change detection method based on spectral feature level fusion and multi-scale segmentation voting decision-making,an algorithm framework is proposed to improve the accuracy of change detection.In this algorithm,the image fusion algorithm is the beginning,the multi-scale features of images are extracted by respectively using different algorithms,and then fused into the feature layer through these extracted features.The change amplitude of the fusion feature vectors between different time phases is measured through Manhattan distance,and the binary change detection chart is obtained after segmentation by the Otsu method.Then,each pixel in the target is marked with the voting strategy in which the minority is subordinate to the majority,finally form the change detection chart.From the final testing results,the high-resolution image classification algorithm based on ridgelet and convolutional neural network is compared with the five most advanced algorithms,and the results have certain advantages.The change detection method based on spectral feature level fusion and multi-scale segmentation voting decision-making is tested in three sets of data.The testing results show that compared with the original spectral features alone and other advanced change detection algorithms,and this method is better.Finally,the algorithm of this paper is applied to the actual data of land survey and has achieved good results,which fully demonstrates the effectiveness of the algorithm research of this paper and its adaptability to actual work.