遥感+深度学习研究综述
Table of Contents
- Deep Learning for Remote Sensing Images
- Traditional Methods for Remote Sensing Images
- Databases
- Survey
Deep Learning for Remote Sensing Images
基于深度学习的遥感图像超分辨率
论文：[TGARS20] Nonpairwise-Trained Cycle Convolutional Neural Network for Single Remote Sensing Image Super-Resolution. [Paper]
代码：https://github.com/haopzhang/CycleCNN
物理方法提升空间分辨率的方法往往会减小传感器进光量以及信噪比，从而导致最终图像画质变差。使用超分辨率算法（super-resolution）则可以绕开前边的物理限制。SR也是一个在图像领域的经典课题，即将低分辨率图像处理为高分辨率图像。
SR方法可以分为两种：
- 单图像超分辨率（single image super-resolution, SISR）
- 多图像超分辨率（multiple images super-resolution, MISR）
SISR更多的应用于遥感领域，因为SISR比起MISR更适用于遥感领域存在多样化的图像传感器的特点。
传统的SR方法主要基于插值和重建。插值方法包括近邻插值，双线性察之，基于深度学习方法的SR目前主要为SRCNN，VDSR和SRRessNet，显示出很好的效果。
基于深度学习的毁坏房屋检测
由于垂直图像的内在限制，以及云层的遮挡，使得其在识别建筑物的结构损伤上存在天然的劣势。
基于深度学习的遥感图像变化检测
S2Looking: A Satellite Side-Looking Dataset for Building Change Detection
这篇工作提出了一个基于卫星图像的建筑物变化监测监督数据集S2Looking。其相较于类似的诸如LEVIR-CD数据集的主要区别为，S2Looking包括同一建筑物的不同拍摄角度，这考验检测模型在角度变化情况下仍然能够识别同一建筑的能力。
包括5000对标注过的图像，大小为1024×1024，0.5～0.8米/像素。将多了的和少了的建筑物都进行了标注。
文章在S2Looking数据集上测试了几个在LEVIR-CD数据集上的SOTA模型，发现结果并不好。分析原因主要有三个：
1) S2Looking有大尺度视野上的观察角变化 2) 明亮度上的差异较大 3) 在农村地区的图像中有各种复杂的场景特征
传统方法很容易被多变的图片细节以及质量区别干扰准确率，因此越来越多的深度学习方法应用到这个领域。包括dual atten- tive fully convolutional Siamese networks (DASNet)， image fusion networks (IFN) ， end-to-end change detection based on UNet++ (CD-UNet++) ， fully convolutional Siamese networks based on concatenation and difference (FC-Siam-Conc and FC-Siam- Diff) ，dual-task constrained deep Siamese convolutional network model (DTCDSCN) 。
DASNet: Dual attentive fully convolutional siamese networks for change detection in high-resolution satellite images
传统的变化检测可以分为两大类：基于像素的变化检测和基于目标的变化检测。传统的基于像素的变化检测往往会忽略空间信息，以及结构语义信息，并且容易受到噪声影响。DASNet模型针对这一问题提出了解决方案，提出了dual attention机制来定位发生变化的区域从而获得更有区分度的特征表示，这使得学习到的特征更加具有鲁棒性。
2020
-
[TGARS20] Remote Sensing Image Super-Resolution Using Novel Dense-Sampling Networks. [Paper]
-
[TGARS20] Remote Sensing Image Super-Resolution via Mixed High-Order Attention Network. [Paper]
-
[TGARS20] Coupled Adversarial Training for Remote Sensing Image Super-Resolution. [Paper]
-
[TGARS20] Nonpairwise-Trained Cycle Convolutional Neural Network for Single Remote Sensing Image Super-Resolution. [Paper]
-
[TGARS20] Improving the Spatial Resolution of Landsat TM/ETM+ Through Fusion With SPOT5 Images via Learning-Based Super-Resolution. [Paper]
-
[TGARS20] Satellite Video Super-Resolution Based on Adaptively Spatiotemporal Neighbors and Nonlocal Similarity Regularization. [Paper]
-
[RemoteSens20] Improved SRGAN for Remote Sensing Image Super-Resolution Across Locations and Sensors. [[Paper]]
2019
- [TGARS19] Super-Resolution of Single Remote Sensing Image Based on Residual Dense Backprojection Networks. [Paper]
Traditional Methods for Remote Sensing Images
2019
- [RemoteSens19] Improving Remote Sensing Image Super-Resolution Mapping Based on the Spatial Attraction Model by Utilizing the Pansharpening Technique. [[Paper]]
2018
- [TGARS18] Adaptive Super-Resolution for Remote Sensing Images Based on Sparse Representation With Global Joint Dictionary Model. [Paper]
Others
Databases
- UC Merced dataset offical
- AID dataset offical
- USC_SIPI image database offical
- NWPU VHR-10 dataset baidu
- NWPU-RESISC45 dataset baidu
