Research On Scene Classification Of Remote Sensing Image Based On Deep Learning

With the rapid development of a variety of earth observation platform systems,such as satellites and unmanned aerial vehicles,it has become increasingly easy to acquire a large number of high-resolution remote sensing images.These high-resolution images have become an important data source for various remote sensing image interpretation tasks.Remote sensing image scene classification is the process of automatically classifying scene images into predefined semantic categories based on their contents.This process must bridge the "semantic gap" between the low-level features of remote sensing images and the more discriminative high-level semantic information.To improve the problem of "semantic gap" and further increase accuracy of remote sensing scene classification,this dissertation uses deep learning to extract more discriminative feature,performs supervised,semi-supervised and unsupervised remote sensing scene classification research as well as super-resolution research of remote sensing image.Extensive experiments on several datasets show that these methods effectively improve the scene classification accuracy.The main work and innovations of this dissertation are as follows:(1)Aiming at remote sensing scene images with diverse ground object categories,complex backgrounds,intraclass diversity and interclass similarity,two more discriminative feature extraction models are proposed to improve the scene classification accuracy.The first is based on a saliency dual-attention residual network model,which embeds spatial attention into the low-level features of the deep residual network to extract cross-spatial salient locations,and embeds channel attention into the high-level features to extract cross-channel salient semantics.The second model uses pre-trained Inception V3 network branches,supervised contrastive learning and gated self-attention modules to extract more discriminative feature representations for improving scene classification accuracy.(2)Aiming at the problem that an insufficient number of labeled remote sensing images are available,a semi-supervised model based on generative adversarial nets is proposed to improve scene classification accuracy.This model introduces a pre-trained external Inception V3 network,a gating unit and a self-attention gated module into the discriminative network to enhance feature representation and thereby improve the semi-supervised classification accuracy.The pre-trained Inception V3 network aims to extract high-level semantic information from remote sensing scene images by fine-tuning;the self-attention gated module is designed to capture long-range dependencies to adaptively focus on important regions in the scene;and the gating unit highlights important features by learning the weight of each feature map and capturing the dependencies between features.(3)For the situation that there are only unlabeled remote sensing scene images in some remote sensing applications,a remote sensing scene classification model based on self-supervised generative adversarial nets is proposed.In this model,similarity loss is introduced to enhance self-supervised representation learning ability;a gated self-attention module is introduced to focus on crucial scene regions and to filter useless background;a pyramidal convolution block is integrated into the residual block of the discriminative networks to enhance feature representation by capturing different levels of details in the image using different types of filters with varying sizes and depths;spectral normalization is introduced into both generative and discriminative networks to stabilize training and enhance feature representations;and multilevel feature fusion is integrated into the discriminative network to achieve more discriminative feature representation.(4)Aiming at the impact of low-and medium-resolution remote sensing images on the accuracy of scene classification,a remote sensing image super-resolution model based on cascade generative adversarial nets is proposed to achieve high-quality super-resolution images at arbitrary multiples.A scene constraint term is integrated into generative network to constrain generated feature to avoid the risk of scene change;content fidelity is introduced into generative network to stabilize the training and avoid vanishing gradient;an edge enhancement module is designed to preserve edge detail and suppress noise;and spectral normalization is introduced into the discriminative network to stabilize the training process.