陈立福, 金昱忱, 李振洪, 宋闯, 王小华, 陈昊达, 刘亚武, 赵俊淇. 基于多特征交叉融合孪生网络的SAR影像地震滑坡识别[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20230479
引用本文: 陈立福, 金昱忱, 李振洪, 宋闯, 王小华, 陈昊达, 刘亚武, 赵俊淇. 基于多特征交叉融合孪生网络的SAR影像地震滑坡识别[J]. 武汉大学学报 ( 信息科学版). DOI: 10.13203/j.whugis20230479
shu
CHEN Lifu, JIN Yuchen, LI Zhenhong, SONG Chuang, WANG Xiaohua, CHEN Haoda, LIU Yawu, ZHAO Junqi. Earthquake-induced Landslides Recognition from SAR Images Based on Multi-feature Cross-fused Siamese Network[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20230479
Citation: CHEN Lifu, JIN Yuchen, LI Zhenhong, SONG Chuang, WANG Xiaohua, CHEN Haoda, LIU Yawu, ZHAO Junqi. Earthquake-induced Landslides Recognition from SAR Images Based on Multi-feature Cross-fused Siamese Network[J]. Geomatics and Information Science of Wuhan University. DOI: 10.13203/j.whugis20230479
shu

基于多特征交叉融合孪生网络的SAR影像地震滑坡识别

Earthquake-induced Landslides Recognition from SAR Images Based on Multi-feature Cross-fused Siamese Network

    摘要
  • 摘要: 由地震引发的群发性大规模地震滑坡是一类非常严重的震后次生灾害，会造成严重的人员伤亡和巨大的经济财产损失。地震发生后迅速准确地识别出这些地震滑坡， 可为国家应急部门评估灾害程度和制定救灾措施提供非常重要的信息指导。 SAR 具有全天时全天候成像特性，但由于地震滑坡背景极为复杂，在 SAR 影像中特征不显著，因此目前识别效果较差。对此，本文提出了一个使用 SAR 图像进行滑坡识别的差异特征与聚合特征交叉融合孪生网络(Difference and Aggregated feature Cross fusion Siamese Network，DACS-Net)。该网络由编码-解码结构组成，编码器采用孪生结构的特征提取网络对地震前后的 SAR 图像进行不同分辨率的特征提取。解码器中通过构建多尺度差异特征和聚合特征生成模块，对不同尺度地震滑坡的特征图进行差异提取与特征聚合，以此来充分表征地震滑坡特性；最后通过提出的多特征交叉融合模块对不同尺度的聚合特征和差异特征进行密集连接逐层解码，提升细节特征与语义特征的提取性能，最终得到识别结果。利用哨兵一号数据对巴布亚新几内亚和米林地震滑坡进行了实验， 独立测试实验的滑坡识别精度为： PA 分别可达 70.75%和 76.5%， Recall 分别可达 60.92%和 71.2%， F1 Score 分别为 65.46%和74.0%， OA 分别为 91.0%和 86.1%。利用该网络，地震滑坡发生后可利用 SAR 影像迅速识别滑坡区，进行灾害应急响应。

     

    Abstract: Objective: Earthquake-induced mass landslides are a severe type of secondary disaster following earthquakes, causing significant casualties and substantial economic losses. Rapidly and accurately identifying these earthquake-induced landslides after an event is crucial for national emergency departments to assess disaster severity and formulate relief measures. Although Synthetic Aperture Radar (SAR) offers all-day, all-weather imaging capabilities, its effectiveness in identifying landslides is currently limited due to the complexity of the background and the less prominent features of landslides in SAR images. Methods: This paper proposes a Siamese network for landslide identification in SAR images, named the Difference and Aggregated feature Cross fusion Siamese Network (DACS-Net). This network, composed of an encoder-decoder structure, utilizes a Siamese feature extraction network in the encoder to extract features of pre- and post-earthquake SAR images at different levels. In the decoder, multiscale difference feature and aggregated feature generation modules are constructed, to perform difference extraction and feature aggregation for landslide feature maps at different scales, fully representing the characteristics of earthquake-induced landslides. The proposed multi-feature cross-fusion module densely connects and decodes different scales of aggregated and difference features layer by layer, enhancing the extraction of detail and semantic features, leading to the identification results. Results: In the experiment, Sentinel-1 data on earthquake-induced landslides in Papua New Guinea and Milin is used, and the results show that the proposed method could effectively identify earthquake-induced landslides. The landslide identification precision (PA) could reach 70.75% and 76.5%, Recall are 60.92% and 71.2%, F1 Score are 65.46% and 74.0%, and Overall Accuracy (OA) are 91.00% and 86.1% for the two cases, respectively. Conclusion: By utilizing this network, achievements have been made in the recognition of landslides in SAR images, enhancing the practical application value of deep learning networks in landslide detection. This holds significant importance for disaster emergency response.

     

    • HTML全文
    • 参考文献(0)
    • 相关文章
    • 施引文献
  • 资源附件(0)

/

返回文章
返回