齐文文,许冲,乔月霞. 2024. 基于谷歌地球引擎和Sentinel-2时序数据的海地多云地区地震滑坡识别. 地震学报,46(4):633−648. DOI: 10.11939/jass.20220168
引用本文: 齐文文,许冲,乔月霞. 2024. 基于谷歌地球引擎和Sentinel-2时序数据的海地多云地区地震滑坡识别. 地震学报,46(4):633−648. DOI: 10.11939/jass.20220168
Qi W W,Xu C,Qiao Y X. 2024. Earthquake-triggered landslides detection in cloudy area of Haiti based on Google Earth Engine and Sentinel-2 time series data. Acta Seismologica Sinica46(4):633−648. DOI: 10.11939/jass.20220168
Citation: Qi W W,Xu C,Qiao Y X. 2024. Earthquake-triggered landslides detection in cloudy area of Haiti based on Google Earth Engine and Sentinel-2 time series data. Acta Seismologica Sinica46(4):633−648. DOI: 10.11939/jass.20220168

基于谷歌地球引擎和Sentinel-2时序数据的海地多云地区地震滑坡识别

Earthquake-triggered landslides detection in cloudy area of Haiti based on Google Earth Engine and Sentinel-2 time series data

  • 摘要: 受云雾影响,强震震后的高分辨率光学影像质量较低,限制了震后地震滑坡调查和评估工作的开展。本文以2021年8月14日海地尼普斯(Nippes) MW7.2地震附近区域为研究区,基于谷歌地球引擎(Google Earth Engine)云平台和Sentinel-2时间序列影像,提出了一种遥感序列影像去云和地震滑坡识别的方法。首先,利用Sentinel-2遥感影像及机器学习算法获取的Sentinel-2云概率数据产品,对长时间序列遥感影像进行去云处理,镶嵌融合得到无云的影像数据;然后对无云光学影像进行多尺度最优分割,利用遥感数据的光谱特征、植被指数、不同类别的空间邻近关系特征等,构建了面向对象的地震滑坡识别模型。结果显示,本文提出的去云和滑坡识别方法有益于多云地区强震震后滑坡空间分布的准确识别,能够为灾后应急调查和评估提供技术支撑。

     

    Abstract: Earthquakes can trigger numerous and widely distributed landslides. Rapid mapping of landslides with high accuracy is crucial for understanding their spatial patterns, and assessing landslide hazards and risks. Visual interpretation or automatic recognition are conventional methods for earthquake-triggered landslide mapping. Due to the influence of post-earthquake rainfall, cloud coverage rate of high-resolution optical remote sensing images is generally high, which limit investigation and assessment of earthquake-triggered landslides. Based on the Google Earth Engine platform, this work proposed a method for removing cloud and cloudless images using Sentinel-2 time series data and the cloud probability dataset. Through comparing performance of different time period, Sentinel-2 time series from 14th August 2021 to 31th January 2022 were chosen with 20% cloud coverage rate. This work masked all pixels covering by cloud and replaced them with cloud-free pixels in the images of adjacent time, producing high quality cloudless mosaic images covering the study area. The proposed method can improve the quality of post-earthquake images effectively. Then, an object-oriented model for earthquake-triggered landslide recognition is constructed based on the cloud-free mosaic images. Firstly, the multiscale segmentation algorithm was applied to the cloudless mosaic images to generate numerous objects. Then, multiple features were calculated based on every object, including spectral features, vegetation index and spatial proximity relationship between adjacent objects belonging to different class. Finally, earthquake-triggered landslides were classified used membership function based on thresholds of multiple features. Taking 2021 Haiti Nippes Earthquake as a case study, this work implemented the proposed method. Meanwhile, evaluation of classification results was carried out based on using manually interpreted landslides using high-resolution Planet images with 3-meter resolution in the validating area. The precision of the method is 77.5%, recall is 52.77% and F1 index is 62.79%. The results show that the proposed methods benefit earthquake-triggered landslides detection in the cloudy area and can provide technical support for post-disaster emergency investigation and assessment.

     

/

返回文章
返回