Research Of Remote Sensing Image Scene Classification Based On Deep Learning

With the rapid development of sensing technology,the data volume of remote sensing images is increasing rapidly,and the resolution is also constantly increasing,which makes the traditional classification methods unable to meet the needs of high-level content interpretation of remote sensing images.As an important research direction in the field of computer vision,scene classification has been widespread concerned and become a hotspot in remote sensing image interpretation.In recent years,convolutional neural network has made many breakthroughs in the field of image processing.It relies on its own characteristics such as parameter sharing,local receptive field,down-sampling and sparse connection.Now it has become a very acti’ve research direction in computer vision and artificial intelligence.Based on the theory of deep learning and combining the characteristics of high-resolution remote sensing images,this thesis discusses how to apply typical convolutional neural network models to remote sensing image scene classification.The main contents of this thesis are as follows:(1)This thesis introduces the existing common remote sensing image scene classification datasets,compares and analyzes the characteristics and problems of these datasets,and explains why to propose new,high-quality,large-scale datasets.After that,it investigates the specific classification strategies of remote sensing image scenes,establishes 31 typical classification scenarios of remote sensing images,and then introduces the specific definition of each category,the source of selection,the factors considered in selection,pixel resolution and so on.Finally,a scene classification dataset with at least 300 images of each category and a total of more than 10,000 remote sensing images is constructed.(2)This thesis explains in detail the basic structure and training process of neural networks.According to the characteristics of remote sensing images and practical training experience,a training method based on transfer learning is selected.According to the actual training effect,continuously adjust and change the structure,learning rate,batch size and other hyperparameters of the convolutional neural network,reduce the over-fitting to obtain the best effect.Finally,we get the excellent effect of convolution neural network model in remote sensing image scene classification.(3)After a series of experiments on different datasets,the training results of different models under different datasets are compared and analyzed.Some results between the datasets and the performance of the training models are obtained,and suggestions on how to train and optimize the convolutional neural network on remote sensing image datasets are given.Finally,we try to apply the trained model to the actual remote sensing image,automate the whole image processing process,and show the visualization results.