毕业生知识库

遥感影像已被广泛地应用于地球资源调查、自然灾害预测和环境监测等诸多任务之中。但由于传感器易受大气密度和云层变化等因素的影响，许多遥感影像存在云层遮挡问题。云层使得图像中获取的地物信息衰减，对进一步的分析造成不利的影响。而对气象分析而言，通过研究云的分布可以发现极端气候现象并总结出相应的变化规律。云检测被看作是遥感图像进行后续识别、分类和分析的关键，是遥感影像修复的重要基础。通过云检测获取云层位置区域信息后，根据近邻地物信息恢复出被遮挡区域，可提升图像的可视化效果；对于遥感地物信息重建，可节省大量卫星资源，避免重复拍摄。另外，对于一些云层较为薄弱的弥散分布区域，如何消除薄云（雾）对可视化的影响和增强地物对比度，对于提升遥感图像的利用率具有同样重要的意义。虽然上述问题在遥感图像领域已经得到广泛的研究，但仍存在准确性偏低和泛化能力偏弱的问题。随着深度学习的兴起，人们尝试将此方法应用于遥感图像，虽然取得了部分研究成果，但仍存在许多关键技术需要突破。

 

为此，本文构建了基于深度学习的面向遥感图像处理的神经网络，以此解决云层分割和云去除中面临的困难。本文的主要工作和贡献如下：
1. 本文提出了一种基于分割模型的卷积神经网络，并将其应用于遥感图像云层区域的提取。在遥感数据方面，本文通过标注大量云层区域构建了一个包含训练、验证、测试部分的数据集。为了解决云层边界模糊导致的误判，在模型结构方面，本文引入边缘检测分支用于约束模型更新并加强云层边界权重。为了解决遥感图像下垫面差异导致的训练发散，本文提出一种由易到难的训练策略：从简单样本训练起并逐渐增大样本难度，使困难样本得到较好的拟合。对比实验表明，本文所提方法可以更快地达到收敛并取得更高云层区域提取精度。
2. 本文设计了一种基于生成式对抗模型的网络结构，并将其应用于遥感图像云去除，通过增强地面对比度、恢复被遮挡区域进而提高遥感图像的可利用率。在数据方面，本文通过手工合成的方式模拟自然场景下的云雾，得到包含训练、验证、测试部分的数据集。结构上，生成网络通过像素级别的监督信息得到去云图像。为了取得更好的去云效果，并且能够利用地面特征的上下文语义信息进行辅助，本文提出使用判别评价图像生成质量的好坏。训练过程中两部分相互博弈，共同提升。实验结果表明本文所提方法相较于自编码机能够取得更加真实的生成效果。

Remote sensing image has been widely applied to many tasks such as earth resources investigation, natural disaster prediction and environmental monitoring. 

However, because of the influence of atmospheric density and cloud change, many remote sensing images are occluded by clouds. 

The cloud layer makes the attenuation of the object information acquired in the image, which has a negative impact on further analysis. 

For meteorological analysis, we can find extreme climate phenomena and summarize the corresponding change rules by studying the distribution of clouds. 

Cloud detection is regarded as the key to remote sensing image recognition, classification and analysis. It's also an important basis for remote sensing image restoration. 

After obtaining the cloud location information through the cloud detection, the image can be improved by restoring the blocked area according to the nearby objects information. 

The reconstruction of remote sensing ground information can save a large number of satellite resources and avoid repeated shooting. 

In addition, for some areas with relatively weak clouds, how to eliminate the influence of the thin cloud (fog) on the visualization and enhance the contrast of the ground objects is of the same significance for improving the utilization of remote sensing images. 

Although the above problems have been extensively studied in remote sensing field, there are still some problems, such as low accuracy and weak generalization ability. 

With the rise of deep learning, people try to apply this method to remote sensing images. 

Although some research results have been obtained, there are still many key technologies needed to break through.
Therefore, for remote sensing image processing, a deep learning based neural network is constructed to solve the difficulties in cloud segmentation and cloud removal. The main work and contribution of this article are as follows:
1. In this paper, a convolutional neural network based segmentation model is proposed and applied to the extraction of cloud area in remote sensing images. 

    In terms of remote sensing data, we set up a dataset containing training, validation and testing part by annotating a large number of cloud regions. 

    In order to solve the misjudgement caused by cloud boundary blurring, the edge detection branch is used to update the model and strengthen the cloud boundary weight in the model structure. 

    In order to solve the training divergence caused by the difference in the surface of remote sensing images, this paper proposes a easy-to-hard training strategy, which is trained from simple samples and gradually increases the difficulty of the sample.

    In this way, the difficult samples can be better fitted. 

    Comparative experiments show that the proposed method can converge faster and achieve higher cloud area extraction accuracy.
2. In this paper, a network structure based on generative adversarial model is designed and applied to remotely sensed image cloud removal. 

    By enhancing ground contrast and restoring the covered area, the utilization rate of remote sensing image can be improved. 

    In terms of data, this paper simulates cloud and fog in natural scenes by manual synthesis, and obtains data sets including training, validation and testing parts. 

    In structure, the generative network obtains cloud images by pixel-level supervision information. 

    In order to obtain better results, this paper proposes to exploit the context semantic information of the ground features. The discriminator is adopted to evaluate the quality of the image generation. In the training process, the two parts are playing together and improving together. 

    Experimental results show that the proposed method is more effective than the auto-encoder machine.