Remote Sensing Image Classification Based On Deep Neural Network And Transitive Transfer Learning

As the representative algorithms of artificial intelligence,deep neural network has the characteristics of high accuracy and no need of artificial feature selection.In recent years,it is widely used in the field of remote sensing image classification.However,compared with traditional machine learning algorithm,deep neural network needs more labeled samples.Transfer learning can use the learned knowledge to help complete new tasks,which can alleviate the above problems to a certain extent.In order to achieve high accuracy classification of remote sensing image with limited labeled samples,this paper combine deep neural network and transfer learning to apply to remote sensing image classification,the specific research contents are as follows,(1)Aiming at the problem that the difference feature between Image Net image and remote sensing image when using Image Net pre-training weight for transfer learning,a remote sensing image classification method based on transfer transfer learning is proposed.This method constructs an intermediate domain dataset with the open source remote sensing scene recognition dataset as the data source,and use Image Net pre-training weight as the source domain,remote sensing images to be classified as the target domain for transfer learning.The connection between Image Net and remote sensing image to be classified is enhanced by adding intermediate domain.While retaining abundant low-level features of Image Net pre-training weight,the effect of transfer learning is improved.(2)Achieve the application of land cover classification in remote sensing images use the proposed transitive transfer learning method.First,the LCC dataset based on open source remote sensing scene recognition dataset is constructed as the intermediate domain,and the ZY-3 image dataset of Panjin area is constructed as the target domain.Then,in order to speed up the update of the weight of the convolutional layer,the fully connected layer of VGG16 is replaced with Global Average Pooling(GAP)layer to designed the GAP-VGG16 network,and the LCC dataset is used for training to obtain the remote sensing image weight.Finally,in order to further extract the remote sensing weight,the convolutional layers are added after the SegNet decoder layer to design T-SegNet,and the obtained weight are transferred to T-SegNet,use ZY-3 Panjin image dataset for training to realize the land cover classification.By selecting different numbers of samples for training,the ZY-3 Panjin image is classified,compared with the traditional deep neural network,the proposed method has higher classification accuracy in the same number of samples.(3)Achieve the application of building extraction in remote sensing images use the proposed transitive transfer learning method.First,the BE dataset based on open source remote sensing scene recognition dataset is constructed as the intermediate domain,and the Inria Aerial Image Labeling Dataset(IAILD)is constructed as the target domain.Secondly,the BE dataset is used to train Image Net pre-training VGG16 to obtain the remote sensing image weight.Then,aiming at the problem of discontinuous extraction of large-scale buildings from the SegNet network,the AS-SegNet network is designed by adding Atrous Spatial Pyramid Pooling(ASPP)module and skip-connection,and the remote sensing image weight is transferred to AS-SegNet,the IAILD dataset is used to train the AS-SegNet.Finally,the building binary classification is performed on the IAILD dataset image to realize the building extraction.By experiments on images of different regions in the IAILD dataset,compared with traditional deep neural networks,the proposed method has higher building extraction accuracy and stronger generalization ability.