Adversarial Learning For Image Classification And Change Detection Using Optical Remote Sensing Images

The rapid development of aerospace technology and sensor technology enables the quick access to massive earth observation data.As one of the most commonly used and accessible remote sensing data,optical remote sensing imagery provides high precision,wide range,and periodic information of earth objects due to its favorable characteristics of high resolution,multi-band of the spectrum,the large image size and the great integrity of the information structure.Being the fundamental and key tasks in the field of analysis and processing for the remote sensing imagery,image classification and change detection have significant application value and strategic importance in many areas of researches,such as military,science and technology,agriculture,environment,natural resources,transportation,urban planning,etc.Therefore,in order to facilitate image classification and change detection for remote sensing imagery,how to extract useful information and features from optical remote sensing imagery quickly and efficiently has become an increasing concern.In recent years,automated analysis and processing have become a prevailing trend in many research fields owing to the rapid development of artificial intelligence(AI)technology and the powerful storage and computing capacity of the current computers.As a supervised learning method,deep learning method has a strong feature learning and expression ability owing to its multi-layer perceptron structure,which realizes the robust automatic model learning by effective using the numerous existing experience and models.This new kind of approach,which differs from conventional AI methods,is been called the deep learning method.And those models used in deep learning methods are called the deep neural network models.After being introduced to the remote sensing field,AI techniques introduce new ideas and provides alternative ways for various traditional remotes sensing tasks.With its powerful expression and fitting ability,the deep learning models have made significant progress in the interpretation and analysis of various remote sensing images.However,for the complex image analysis task of image classification and change detection,deep learning methods also has certain limitations.With the growing depth of the hierarchical structure of a deep neural network,the capacity of this network has exceeded the amount of information embedded in the images.This often leads to the weak generalization ability of the existing deep neural networks.This means,rather than learning how to solve a problem essentially,deep neural networks mechanically record the end-to-end mapping rules.In the tasks of image classification and change detection,most deep learning methods only show excellent performance in the training data set.Once new data is added,the accuracy of the existing trained model will be extremely reduced.Aiming at the problem of intelligent processing of optical remote sensing imagery,this paper carries out in-depth research in combine with the advantages and existing problems of deep neural network.The main research contents and innovations include:1)This paper systematically summarizes various deep learning methods currently used for optical remote sensing image classification and change detection.Based on the characteristics and difficulties of image classification and change detection tasks,this paper discusses and studies the current commonly used deep learning methods.2)The superiority of deep adversarial learning in image classification and change detection is elaborated in detail.Based on the current problems of intelligent remote sensing image processing and the fundamental theories of deep learning and deep adversarial learning,this paper analyzes the applicability,feasibility,and advantages of deep adversarial learning method used for image classification and change detection for optical remote sensing imagery in-depth.3)This paper proposes a multi-model joint framework,where two self-attention module are embedded in a Pix2 pixGAN,for image classification on homogeneous remote sensing images.Combined with the basic process and related technologies of homogeneous image classification,this paper designs an embedded adversarial network to realize this task by multiple constraints.This method effectively solves the overfitting problem of traditional deep learning methods and achieves to obtain higher image classification accuracy by using a smaller deep learning model.4)This paper proposes a multi-model joint framework,where a geometryconsistent domain adaptation module is embedded in a co-training adversarial network,for image classification on heterogeneous remote sensing images.Traditional deep learning methods require a large amount of annotated data for model training.However,labeling complex optical remote sensing images is time-consuming and labor-extensive.For this problem,this paper constructs a domain adaptation-based co-training adversarial network,and realizes heterogeneous image classification by using transfer learning technology.This method enables learning knowledge from labeled paired data and transfers it to unlabeled data to realize unsupervised image classification.5)This paper proposes a multi-model joint framework,where a conditional dual learning framework and a Siamese neural network are paralleled,for change detection on bi-temporal remote sensing images by the parallel neural network training strategy.Due to the variance of environmental conditions for bi-temporal images,the domain difference between the bi-temporal images severely influences the accuracy of change detection results.For this problem,this paper constructs a dual learning-based Siamese framework,and achieves change detection by two parallel streams.This proposed method effectively reduces the negative impact of domain difference on the change detection task and quickly realize the pixel-level image change detection.