测绘学报 ›› 2026, Vol. 55 ›› Issue (10): 1839-1851.doi: 10.11947/j.AGCS.2026.20260095

• 摄影测量学与遥感 • 上一篇    下一篇

DC-LightSeg:基于可变形卷积的轻量化滑坡语义分割模型

刘瑞1,2(), 姚东辰1,2(), 周棋1,2   

  1. 1.成都理工大学地球物理学院,四川 成都 610059
    2.成都理工大学地质灾害防治与地质环境保护全国重点实验室,四川 成都 610059
  • 收稿日期:2026-03-19 修回日期:2026-08-27 出版日期:2026-11-16 发布日期:2026-11-16
  • 通讯作者: 姚东辰 E-mail:lr@cdut.edu.cn;yaodongchen@stu.cdut.edu.cn
  • 作者简介:刘瑞(1979—),男,博士,副教授,研究方向为地球探测与信息技术。E-mail:lr@cdut.edu.cn
  • 基金资助:
    地质灾害防治与地质环境保护全国重点实验室开放基金(SKLGP2022K026);四川省教育厅人文社会科学重点研究项目(ZHYJ21-YB04);福建省空间信息感知与智能处理重点实验室项目(FKLSIPIP1002)

DC-LightSeg: a lightweight landslide semantic segmentation model based on deformable convolutions

Rui Liu1,2(), Dongchen Yao1,2(), Qi Zhou1,2   

  1. 1.College of Geophysics, Chengdu University of Technology, Chengdu 610059, China
    2.State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, Chengdu 610059, China
  • Received:2026-03-19 Revised:2026-08-27 Online:2026-11-16 Published:2026-11-16
  • Contact: Dongchen Yao E-mail:lr@cdut.edu.cn;yaodongchen@stu.cdut.edu.cn
  • About author:Liu Rui (1979—), male, PhD, associate professor, majors in geodetection and information technology. E-mail: lr@cdut.edu.cn
  • Supported by:
    Opening Fund of State Key Laboratory of Geohazard Prevention and Geoenvironment Protection(SKLGP2022K026);Key Research Project of Humanities and Social Sciences of Sichuan Provincial Department of Education(ZHYJ21-YB04);Fujian Key Laboratory of Spatial Information Perception and Intelligent Processing(FKLSIPIP1002)

摘要:

利用遥感影像实现准确、高效的滑坡地质灾害识别对灾害评估与应急管理具有重要意义。然而,滑坡的形态不规则、尺度差异显著且背景复杂,现有方法难以兼顾分割精度与计算效率。针对上述问题,本文提出一种基于可变形卷积的轻量化语义分割模型DC-LightSeg。该模型结合大核分组卷积与可变形卷积动态调整感受野,精准捕捉滑坡的不规则边界和破碎特征;同时,引入软池化(SoftPool)与改进的双向特征金字塔网络(BiFPN),以加强小尺度滑坡特征保留和多尺度信息融合。在毕节和尼泊尔两个公开滑坡数据集上的试验结果表明,DC-LightSeg参数量仅为1.23×106,并取得较有竞争力的分割性能:毕节数据集上的F1值、召回率、精确率和交并比(IoU)分别为88.78%、88.96%、88.60%和79.82%;尼泊尔数据集上的F1值、召回率、精确率和交并比分别为72.06%、78.40%、66.67%和56.33%。结果表明,该模型在较低参数量下兼顾了识别精度与计算效率,可为资源受限条件下的滑坡遥感快速识别提供参考。

关键词: 滑坡识别, 语义分割, 可变形卷积, 轻量化, 双向特征金字塔

Abstract:

Landslides, as a serious geological hazard, are of great significance for disaster assessment and emergency management through accurate and efficient landslide identification using remote sensing images. However, landslides have irregular shapes, significant scale differences, and complex backgrounds, making it difficult for existing methods to balance segmentation accuracy and computational efficiency. In response to the above issues, this article proposes a lightweight semantic segmentation model DC-LightSeg based on deformable convolution. This model dynamically adjusts the receptive field by combining large kernel grouped convolution and deformable convolution, accurately capturing the irregular boundaries and fragmentation features of landslides; Meanwhile, it introduces SoftPool and an improved bidirectional feature pyramid network (BiFPN) to effectively enhance the preservation of small-scale landslide features and multi-scale information fusion. The experimental results on two publicly available landslide datasets, Bijie and Nepal, showed that the DC-LightSeg has only 1.23×106 parameters and achieved competitive segmentation performance: the F1score, Recall, Precision, and intersection over union (IoU) on the Bijie dataset were 88.78%, 88.96%, 88.60%, and 79.82%, respectively; On the Nepal dataset, they are 72.06%, 78.40%, 66.67%, and 56.33%, respectively. The results indicate that the model balances recognition accuracy and computational efficiency at a lower parameter count, and can provide reference for rapid landslide remote sensing identification under resource limited conditions.

Key words: landslide recognition, semantic segmentation, deformable convolution, lightweighting, bidirectional feature pyramid

中图分类号: