基于光学遥感影像的区域滑坡普查易受云雾天气的影响，存在滑坡体调查不全面的问题，无法满足震后应急调查与恢复重建的需求。本文提出了一种极化SAR卫星数据滑坡普查方法，采用高分三号全极化SAR卫星影像数据，以九寨沟地震震区为实验区，在深入分析滑坡体和其他地物类型散射特征的基础上，融合极化特征、纹理特征和地形特征等多维特征信息，结合高分二号影像获取的训练样本，构建基于BP神经网络的全极化SAR数据滑坡自动识别模型，实现滑坡体的自动快速识别。与高分辨率光学影像与无人机航空影像目视解译结果相比较，总体识别精度为92.8%，Kappa系数为0.715，识别准确度满足地震应急实际应用的需求。研究成果可用于震区大区域滑坡体的普查，为后续开展无人机高分辨率影像滑坡体详查、灾后应急与景区恢复提供辅助信息支撑，并促进国产高分SAR卫星数据在防震减灾中的应用。

Conventional landslide identification model is only suitable for multispectral images. Optical images can directly reflect the information of ground objects and provide people a real and intuitive feeling. Earthquake areas are often hit with inclement weather after the occurrence of earthquakes.
Effective optical images are difficult to obtain in the presence of clouds and fog, thus leading to the incomplete recognition of landslide information. Given that SAR satellite technology is not affected by clouds and breaks through the limitations of optics, SAR data have gradually become the mainstream data for earthquake disaster response and assessment. With the development of SAR sensors, SAR has developed into a multiband, multi-polarization, multi-angle, and variable working model from the first single mode of operation. SAR data guarantee flexibility in its application to seismic landslide identification. Existing SAR image landslide identification methods mainly use single characteristics of texture features or polarization features of landslide information in SAR images. Multi-source features in SAR images have not been fused, especially the characteristics of multi-polarization SAR image data. Thus, landslide survey accuracy is low and cannot meet actual application needs.
Taking the Jiuzhaigou earthquake as an example, this study adopts the first C-band multi-polarization GF-3 satellite data with a resolution of 1 m as the data source. Polarization and texture features of the image are extracted based on an in-depth analysis of the characteristics of multi-polarization image data. Afterward, GF-2 satellite data obtained post-earthquake are carefully registered with the GF-3 satellite data. Typical landslide samples are selected from the GF-2 images, which are used as training samples for classification. Finally, back propagation neural network is used to extract the landslide in the whole area through the comprehensive utilization of polarization characteristics, texture, and terrain feature information based on the training samples. To meet the urgency of the disaster, identification accuracy can be improved as much as possible while meeting the efficiency of information identification. The findings can provide a reference for the restoration, reconstruction, and scientific exploration of the Jiuzhaigou earthquake.
A comparison of the results of visual interpretation of GF-2 optical images and the unmanned aerial vehicle images revealed that the overall extraction accuracy of the landslide is 92.8% and the Kappa coefficient is 0.715. The scattering characteristics of the ridge on the image are easily confused with the bright spots formed by the landslide because these characteristics are more obvious. The spatial distribution feature of the slope can eliminate partial landslide information and eliminate the influence of the ridge.
Taking the Jiuzhaigou earthquake as an example and using the homemade GF-3 full polarization SAR satellite data, this study proposes a fully polarimetric data seismic landslide automatic recognition method based on integrated polarization features, texture features, and terrain features. The method is used for a general investigation of landslides in the entire earthquake area of Jiuzhaigou. The extraction results meet the requirements of earthquake emergency, post-earthquake recovery, and reconstruction. The method also promotes the application of GaoFen satellites in the earthquake prevention and disaster reduction industry.