遥感影像场景分类综述

Remote Sensing Image Scene Classification Using CNN-CapsNet

Attention-Based Deep Feature Fusion for the Scene Classification of High-Resolution Remote Sensing Images

Remote Sensing Scene Classification by Gated Bidirectional Network

Robust space–frequency joint representation for remote sensing image scene classification

Scale-free convolutional neural network for remote sensing scene classification

Graphical representation of the architecture of (a) original model, (b) our SF-CNN model, and © FCL convolution.

A feature aggregation convolutional neural network for remote sensing scene classification

Overall structure of the proposed FACNN. FACNN is mainly composed of three components: backbone pipeline, convolutional feature encoding module, and softmax classifier. The convolutional feature encoding module, as shown in the orange box, is proposed to encoding the intermediate convolutional features into the convolutional representation through the end-to-end training. First, intermediate convolutional features are fed to the pooling layer to unify the width and height of different features to 7×7 . Second, the concatenation layer is used to concatenate different features by the channel. Third, the 1×1 convolutional layers with ReLU are utilized to fuse the information of different features among the channel. Then, the global average pooling layer is employed to generate the convolutional representation. The convolutional representation is merged with the second FC feature to generate the discriminative scene representation. Finally, the discriminative scene representation is fed to the softmax classifier to obtain the semantic labels of remote sensing scenes.

Scene classification using hierarchical Wasserstein CNN

Pipeline of our proposed method. The pipeline primarily involves the following steps: 1) fine-tuning three pretrained CNNs using RS training sets; 2) extracting the validation images’ deep features from three fine-tuned CNNs and concatenating them; 3) building a category tree using the concatenated features; 4) training individual HW-CNNs; and 5) conducting testing.

Scene classification based on multiscale convolutional neural network

When deep learning meets metric learning: Remote sensing image scene classification via learning discriminative CNNs

Illustration of the core idea of the proposed D-CNN method. To address the challenges of within-class diversity and between-class similarity, we propose to learn D-CNNs by optimizing a new objective function. Apart from minimizing the cross-entropy loss, we also impose a metric learning regularization term on the CNN features to enforce the D-CNN models to be more discriminative. Thus, in the D-CNN feature spaces, the images from the same scene class are as close as possible and the images of different classes are as far away as possible.

Remote sensing scene classification using multilayer stacked covariance pooling

Flowchart of the proposed MSCP classification framework. The proposed framework consists of three steps: 1) multilayer feature extraction using a pretrained CNN model; 2) MSCP; and 3) SVM classification. Dotted and colored lines: downsampling and channelwise average fusion, respectively.

Fusion High-and-Low-Level Features via Ridgelet and Convolutional Neural Networks for Very High-Resolution Remote Sensing Imagery Classification

Siamese Convolutional Neural Networks for Remote Sensing Scene Classification

Attention Consistent Network for Remote Sensing Scene Classification

Multilayer Feature Fusion With Weight Adjustment Based on a Convolutional Neural Network for Remote Sensing Scene Classification

Framework of MLFF-WA.

Crossing access with weight adjustment.

A Multi-Level Convolution Pyramid Semantic Fusion Framework for High-Resolution Remote Sensing Image Scene Classification and Annotation

Self-Attention-Based Deep Feature Fusion for Remote Sensing Scene Classification

Multilayer Feature Fusion Network for Scene Classification in Remote Sensing