Design And Realiization Of Optical Remote Sensing Image Target Recognition Based On Convolutional Neural Network

With the development of remote sensing technology,the image datasets of sub-meter spatial images acquired nowadays are increasing day by day.There is also a large amount of unexplored information in these high-resolution remote sensing images.This dissertation discusses the application of deep convolutional neural network to UCMerced-Landuse dataset,and proposes to use UCMerced-Landuse dataset to train and transfer the learning characteristics of convolutional neural network from natural image recognition tasks to High resolution optical remote sensing image recognition task.In the latest technological and commercial applications,a large number of VHR image datasets have been accumulated in the visible RGB spectral bands,and this work explores the potential of deep learning algorithms that can be used to automate high-resolution optical remote sensing image datasets UCMerced-Landuse for classification.The benefits of automatic visible band VHR UCMerced-Landuse classification include applications such as automatic change detection or mapping.Recent work has demonstrated the potential of deep learning methods for land-use classification;however,this paper improves upon existing techniques by extending existing data sets using eligible geospatial data.In addition,the generalization of classifiers was tested by extensively evaluating the classifiers on the unknown dataset,and we counted the accuracy of each classifier,and the actual results exceeded the benchmarks used in the training.Deep networks have many parameters,so they are usually built using a very large set of tag data.Large datasets suitable for UCMerced-Landuse are not readily available,but Refinement Learning allows retraining of networks that train one task to perform another recognition task.The contributions of this dissertation include: Prove that deep networks trained for image recognition in one task can be effectively migrated to remote sensing applications and perform better than hand-labeled classifiers without the need for extensive training data sets.This was confirmed on the UC Merced dataset,where a mean accuracy of 97.66% was achieved using convolutional neural networks and migration learning.These results are further confirmed by the uncorrelated VHR images at the same resolution as the training set.