适于巨灾风险暴露数据库构建的双通道多任务建筑空间与用途识别算法

A dual-channel multi-task building space and usage recognition algorithm suitable for catastrophe risk exposure database construction

摘要

为有效应对金融保险业在巨灾风险精细化管理实践中构建巨灾风险暴露数据库环节所面临的我国建筑空间及用途特征高精度识别挑战，本文提出了一种基于双通道多任务架构的建筑空间与用途识别算法。算法采用双通道输入设计，融合结构化社会属性数据与多波段遥感影像，在编码器中引入残差结构以增强深度网络训练稳定性，解码器则采用带跳跃连接的设计，实现多层次特征的有效融合。算法整体架构为多任务结构，浅层网络任务用于提取建筑物边缘和纹理信息，深层网络任务用于识别建筑物用途，通过参数和特征共享机制提升识别精度。算法基于全国90个城市中采集的10万栋高质量标注数据进行训练，在测试集中表现出良好的泛化能力。模型在建筑覆盖区域(约占全国面积54%)的3.51亿个高分辨率样本上完成推理，生成了高精度的建筑空间与用途识别结果。本文的双通道多任务算法显著提升了巨灾风险暴露数据库的核心特征识别精度，其大规模推理结果可直接服务于保险行业标的级风险评估业务，为金融保险业依托巨灾模型实施精细化风险管理提供了关键技术支撑。

Abstract

To effectively address the challenge of high-precision identification of building space and usage characteristics in the construction of a catastrophe risk exposure database during the refined management of catastrophe risks in the financial and insurance industry, this paper proposes a building space and usage identification algorithm based on a dual-channel, multi-task architecture. The algorithm employs a dual-channel input design, integrating structured social attribute data and multi-band remote sensing imagery. A residual structure is introduced into the encoder to enhance the training stability of the deep network, while the decoder uses a design with skip connections to achieve effective fusion of multi-level features. The overall algorithm architecture is a multi-task structure and the shallow network tasks are used to extract building edge and texture information, while deep network tasks are used to identify building uses. Parameter and feature sharing mechanisms improve identification accuracy. The algorithm is trained on high-quality labeled data from 100 000 buildings collected in 90 cities nationwide, demonstrating good generalization ability on the test set. The model completes inference on 351 million high-resolution samples covering approximately 54% of the national building area, generating high-precision building space and usage identification results. The dual-channel multi-task algorithm presented in this paper significantly improves the accuracy of core feature identification in the catastrophe risk exposure database. And its large-scale inference results can directly serve the object-level risk assessment business of the insurance industry, providing key technical support for the financial and insurance industry to implement refined risk management based on catastrophe models.

关键词

巨灾保险 / 巨灾模型 / 人工智能 / 风险暴露数据库 / 多源异构 / 多任务技术

Key words

catastrophe insurance / catastrophe model / artificial intelligence / risk exposure database / multi-source heterogeneous / multi-task technology

参考文献

[1] 史培军, 李宁, 刘婧, 等 . 探索发展与减灾协调之路: 从2006年达沃斯国际减灾会议看中国发展与减灾协调对策[J]. 自然灾害学报, 2006, 15(6): 1-8.

[2] SHI Peijun, LI Ning, LIU Jing, et al. Discussion on harmonization way of development and disaster reduction: Harmonization strategies of development and disaster reduction in China based on “The International Disaster Reduction Conference, Davos, 2006”[J]. Journal of Natural Disasters, 2006, 15(6): 1-8. (in Chinese)

[3] 史培军, 杜鹃, 叶涛, 等 . 加强综合灾害风险研究, 提高应对灾害风险能力: 从第6届国际综合灾害风险管理论坛看我国的综合减灾[J]. 自然灾害学报, 2006, 15(5): 1-6.

[4] SHI Peijun, DU Juan, YE Tao, et al. Enhancing researches on integrated disaster risk and improving disaster risk coping capacity of business and industry: Integrated disaster reduction in China based on “the Sixth Annual IIASA-DPRI Forum on Integrated Disaster Risk Management, 2006”[J]. Journal of Natural Disasters, 2006, 15(5): 1-6. (in Chinese)

[5] 曾立新, 孙立娟, 吕文栋. 巨灾保险下减灾积极性的影响因素及激励措施: 基于美国巨灾保险项目的经验[J]. 自然灾害学报, 2011, 20(3): 38-46.

[6] ZENG Lixin, SUN Lijuan, LV Wendong. Influence factors and encouraging measures of initiative in disaster reduction under catastrophic insurance: Experience based on catastrophic insurance programs in the United States[J]. Journal of Natural Disasters, 2011, 20(3): 38-46. (in Chinese)

[7] 李冲, 朱平安, 王慧彦. 我国建立巨灾保险制度探析[J]. 自然灾害学报, 2010, 19(2): 47-53.

[8] LI Chong, ZHU Ping′an, WANG Huiyan. Exploration and analysis of constructing catastrophe insurance in China[J]. Journal of Natural Disasters, 2010, 19(2): 47-53. (in Chinese)

[9] Independent group of scientists appointed by the secretary-general. Global sustainable development report 2019: The future is now-science for achieving sustainable development[R]. New York: United Nations, 2019.

[10] 张攀. 风险暴露数据提取方法及对地震保险费率厘定的影响研究[D]. 北京: 中国地震局地球物理研究所, 2023.

[11] ZHANG Pan. Research on risk exposure data extraction methodlolgy and impact on earthquake insurance rates[D]. Beijing: Institute of Geophysics China Earthquake Administration, 2023. (in Chinese)

[12] WANG M Z, SU L L, CI L Y, et al. RSBuilding: Towards general remote sensing image building extraction and change detection with foundation model[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 3439395.

[13] SUN X, WANG P J, LU W X, et al. RingMo: A remote sensing foundation model with masked image modeling[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5612822.

[14] CHEN K Y, LIU C Y, CHEN H, et al. RSPrompter: Learning to prompt for remote sensing instance segmentation based on visual foundation model[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 3356074.

[15] LI Z H, HE W, LI J P, et al. Identifying every building’s function in large-scale urban areas with multi-modality remote-sensing data[C]// 2024 IEEE International Geoscience and Remote Sensing Symposium. Piscataway: IEEE, 2024: 310-314.

[16] LI L W, ZHU J M, GAO L R, et al. Detecting and analyzing the increase of high-rising buildings to monitor the dynamic of the Xiong'an New Area[J]. Sustainability, 2020, 12(11): 4355.

[17] YIN J D, DONG J W, HAMM N A S, et al. Integrating remote sensing and geospatial big data for urban land use mapping: A review[J]. International Journal of Applied Earth Observation and Geoinformation, 2021, 103: 102514.

[18] GUO X, LAO J W, DANG B, et al. SkySense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2024: 27662-27673.

[19] KONG B, AI T H, ZOU X Y, et al. A graph-based neural network approach to integrate multi-source data for urban building function classification[J]. Computers, Environment and Urban Systems, 2024, 110: 102094.

[20] ZHOU W, PERSELLO C, STEIN A. Building usage classification using a transformer-based multimodal deep learning method[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 16: 4567-4580.

[21] ZHANG Z X, QIAN Z, ZHONG T, et al. Vectorized rooftop area data for 90 cities in China[J]. Scientific Data, 2022, 9: 4057-4076.

[22] GB/T 35648- 2017 地理信息兴趣点分类与编码[S]. 北京: 中国标准出版社, 2017.

[23] GB/T 35648- 2017 Classification and coding of geographic information points of interest[S]. Beijing: Standards Press of China, 2017. (in Chinese)

[24] LU W P, TAO C, LI H F, et al. A unified deep learning framework for urban functional zone extraction based on multi-source heterogeneous data[J]. Remote Sensing of Environment, 2022, 270: 112830.

[25] 黄晓鸣, 何富运, 唐晓虎, 等 . U-Net及其变体在医学图像分割中的应用研究综述[J]. 中国生物医学工程学报, 2022, 41(5): 567-576.

[26] HUANG Xiaoming, HE Fuyun, TANG Xiaohu, et al. Review on applications of U-Net and its variants in medical image segmentation[J]. Chinese Journal of Biomedical Engineering, 2022, 41(5): 567-576. (in Chinese)

[27] HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA: IEEE, 2016: 770-778.

基金资助

国家重点研发计划项目(2023YFC3008500)