Research On Key Techniques Of Object-Oriented Remote Sensing Pattern Recognition Based On Deep Neural Network

Object-oriented remote sensing technology has been introduced into the field of remote sensing since 2000,providing better decision support for urban land optimization and smart city construction.In recent years,with the availability of high spatial resolution remote sensing images such as GF-2,it has become more important to understand how to better mine the ground features contained in the images.With the deepening of object-oriented remote sensing technology,the defects of the traditional object-oriented remote sensing pattern recognition technology are gradually emerging,which obstructs the application of object-oriented object-oriented remote sensing technology in actual production and life,Based on deep neural network,this paper extends and extends the traditional object-oriented remote sensing pattern recognition technology.This thesis mainly contains the following three pointsThe first is the establishment of a basic framework for object-oriented remote sensing classification based on convolutional neural networks:(1)An object-oriented remote sensing data method with the advantages of objects and pixels is designed for convolutional neural networks(2)The basic flow of object-oriented remote sensing classification based on product neural network have been introduced(3)Combined with perceptual hashing algorithm,remote sensing image object feature map extracted by convolutional neural network have been analyzed The feasibility of applying convolutional neural networks to object-oriented remote sensing is demonstrated.The second is the parallel deep neural networks classification based on spectralspatial information:(1)According to the characteristics of high spatial resolution remote sensing data,two different convolutional neural networks are designed to extract spatial features and spectral features and classify them;2)A spectral-spatial information parallel convolution neural network classification framework have been established based on the feature-level ensemble learning perspective;(3)The reason for the upgraded precision compared with single branch was finally analyzed by methods such as complementarity analysis,classification feature visualization,and target coarse-grained output.The second is a multi-scale remote sensing classification method based on threedimensional deep convolution strategy is proposed: 1)From the perspective of deep three-dimensional convolutional neural networks,an object-oriented remote sensing data scaling-up dataset is established;2)taking the high-level remote sensing object class inference as the task,based on deep 3D convolutional neural network algorithm,adopting a data-driven approach to the mining of scaling-up rules,and eventually high-level remote sensing objects are classified(3)From perspective of undermining the area-dominated,importance-dominated rules,its advantages over two-dimensional convolutional neural networks has been analyzed.This thesis combines object-oriented remote sensing technology and deep neural network to solve the problems of feature design and feature selection in traditional object-oriented remote sensing pattern recognition.the two tasks including object remote sensing classification and multi-scale remote sensing classification have been solved.The key technologies involved were discussed to better facilitate the application of deep neural networks in the field of object-oriented remote sensing pattern recognition,thus making its own contribution to decision support in related fields.