Data Augmentation Method For Scene Classification Of High-resolution Remote Sensing Images

The feature extraction of remote sensing images is a basic and important task in the field of remote sensing.How to accurately detect,classify and map objects in remote sensing images has extremely important significance and applications for change detection,geographic data update,disaster assessment,etc.The scene classification is one of the hot research directions in the field of remote sensing.It is an important way to resolve remote sensing image information,which is of great significance to the understanding of the real world.The deep learning method,which can automatically extract the feature expression through the multi-layer network,is mainly to extract the high-level semantic information of the images and does not require lots of engineering skills and domain expert knowledge.At this stage,deep learning methods have become popular in the field of remote sensing.The method based on deep learning needs a large number of samples to guide the neural network model to extraction the image feature and consumes too many computing resources.However,the production of a large number of high-quality remote sensing image training samples requires a lot of time and labor costs and is limited by time and space.The data augmentation method is an important way,which can greatly reduce the time required for manual labeling and improve the accuracy of feature extraction.To solve the shortcomings of existing data augmentation methods,an improved sample enhancement method was proposed and the NWPU remote sensing image dataset was used as experimental data.The main research content and corresponding results of this paper are as follows:(1)In this paper,the class activation mapping mechanism(Grad-CAM)was used to combine the traditional data augmentation methods to propose a semi-supervised data augmentation method.And the images in the remote sensing dataset were amplified by new methods in the experiment.The proposed method in this paper can prevent CNNs network from learning wrong label information,and increase the overall classification accuracy of remote sensing image dataset scene classification by more than 0.4%.(2)In this paper,a new generative adversarial network-Rel-SAGAN-was proposed by integrating the self-attention mechanism and the loss function of Relativistic GAN,(RGAN),and this model has a bigger kernel size and a deeper convolutional layer network.The proposed model synthesizes higher-quality remote sensing natural scene data samples.And the overall classification accuracy of remote sensing datasets combined with high-quality synthetic samples evaluated by FID has been improved by more than 3%,which can further alleviate the problem of the lack of samples in remote sensing datasets.