Research On Remote Sensing Image Data Augmentation Based On Adversarial Learning

With the development of remote sensing technology,remote sensing images have been widely used in the fields of change detection and remote sensing interpretation.However,problems such as small number,poor quality and insufficient diversity of remote sensing images limit the performance improvement of remote sensing interpretation and other follow-up studies.Numerous studies have shown that data augmentation is an effective means to solve above problems.By learning the data distribution of training samples,the method can synthesize realistic remote sensing images and achieve the purpose of data expansion.Accordingly,in this paper,two image generation models for remote sensing data augmentation based on generative adversarial nets are proposed.These two models can generate high quality images with rich variations and greatly shorten the generation time,which can provide data support for algorithms such as remote sensing interpretation.The traditional data augmentation algorithm is based on the basic principle of image processing and carries out geometric transformation or random perturbation of pixel value on the original image.Although the number of images is expanded,the corresponding image labels cannot be obtained and the quality of the generated images is poor.In recent years,data augmentation algorithms based on neural networks have improved the quality of generated images,but the generation time is long and the generated images are single.In addition,few studies applied to the field of remote sensing images.In view of the shortcomings of existing data augmentation algorithms,this paper proposes two generation models based on China’s remote sensing images.The main work of this paper is as follows.1)A deeply supervised generative adversarial model(D-s GAN)is proposed to generate remote sensing images.Since most of the deep learning algorithms in the remote sensing field belong to the category of supervised learning,it is necessary to generate labels corresponding to images.At the same time,higher requirements are put forward for the quality and speed of generated images.Taking the above factors into consideration,the fusion result of "segmentation image + random noise" is taken as the input in this paper to generate labels at the same time as the image.The generator structure of the model adopts UNET ++ and the down-sampling process innovatively introduces the segmentation image of remote sensing images,which reduces the semantic loss of images and improves the quality of generated images.In order to reduce the time of image generation,multiple discriminators are used to monitor the subnetwork of the generator.The experimental results show that compared with Co GAN,Sim GAN and Cycle GAN,this model has better performance on FCN-scores and shorter generation time.In addition,the introduction of image segmentation makes the generated samples with corresponding labels,which can meet the needs of supervised learning.2)A basing text deeply supervised generative adversarial model(BTD-s GAN)is proposed to generate remote sensing images.Aiming at the problem of insufficient diversity of remote sensing images generated by data augmentation algorithm,this paper introduces text feature constraints in the process of image generation and constructs the BTD-s GAN model.The specific work is as follows.First,the generator uses Perlin noise and encoded text feature vector as input.Second,the down-sampling process of generator and discriminant use text feature as constraint.Third,the generator’s subnetwork is supervised by two discriminants with the same structure.Results of experiments show that compared with the current generation model based on text description,this model can achieve faster generation speed and generate more abundant images.This paper constructs the above two models from different perspectives.D-s GAN solves the problems of poor image quality,insufficient image labels and slow image generation.BTD-s GAN solves the problem of insufficient diversity of generated images.In order to test the effect of the model,this paper generated images based on the remote sensing data of Anhui and Jiangxi provinces of China and used them in the remote sensing interpretation project of the Ministry of Water Resources of China.The final results show that the remote sensing data generated by the model helps improve the accuracy of the interpretation network by 9%,and the model proposed in this paper meets the actual generation requirements.