摘要
为探索资源三号03星立体像对在地形级实景三维数据更新中的应用潜力,本文以广东省河源市龙川县多地形区域为试验区,提出融合ResNet-18改进网络与时空注意力机制的深度学习特征匹配方法,结合区域网平差技术优化影像几何关系,并构建数字表面模型(DSM)与正射影像(DOM)。通过与传统RPC+有理多项式模型及高分七号、WorldView-3卫星数据对比分析,结果表明:深度学习框架使山地场景特征点匹配Dice系数提升31.6%,平面与高程中误差分别降低32.6%和24.8%,丘陵及山地场景误差增长斜率减少28.5%,一定程度上可以突破卫星原始分辨率限制,验证了智能化算法在复杂地形中的强适应性,研究可为国产卫星通过技术融合实现高精度、低成本广域监测提供有效参考。
To explore the application potential of ZY-303 satellite stereo pairs in terrain-level real-scene 3D data updating,this study selected the multi-terrain area of Longchuan County,Heyuan City,Guangdong Province as the test area.A deep learning feature matching method integrating an improved ResNet-18 network with a spatiotemporal attention mechanism was proposed,combined with block adjustment techniques to optimize image geometric relationships and construct digital surface models(DSM)and digital orthophoto maps(DOM).Comparative analyses with traditional RPC+rational polynomial models,GF-7,and WorldView-3 satellite data revealed that the deep learning framework improved the Dice coefficient of feature point matching in mountainous areas by 31.6%,reduced mean plane and elevation errors by 32.6%and 24.8%,respectively,and decreased error growth slopes in hilly and mountainous terrains by 28.5%,demonstrating its ability to partially overcome satellite resolution limitations and strong adaptability to complex terrain.This study provides an effective reference for domestic satellites to achieve high-precision,low-cost wide-area monitoring through technological integration.
作者
陈子莹
CHEN Ziying(Guangzhou Urban Planning&Design Survey Research Institute Co.,Ltd.,Guangzhou 510000,China)
出处
《城市勘测》
2025年第4期78-82,共5页
Urban Geotechnical Investigation & Surveying
作者简介
第一作者:陈子莹(1988-),女,城乡规划工程师,主要从事测绘、城乡规划管理工作。E-mail:Ceitghwody@163.com。
