摘要
对于当前利用深度学习技术实现遥感影像中道路提取时,由于网络中的矩形卷积而导致无法捕捉道路形态特征、提取精度低等问题。基于此,提出一种基于形变卷积的卷积神经网络(D-Net,Deformation Network),该网络旨在从遥感影像中准确提取道路区域。为了验证模型的性能,在马萨诸塞州道路数据集和DeepGlobe道路数据集上进行了测试。此外,还选择了一景GF-2影像对模型的泛化能力进行评估。实验结果显示,D-Net显著提升了道路分割的精度,为通过形变卷积优化提取结果提供了有力支持。该研究的开展对于进一步探索道路智能提取具有重要的理论和实践指导意义。
In the current road extraction from remote sensing images using deep learning techniques,the issue arises due to the use of rectangular convolutions in the network,leading to difficulties in capturing road morphology features and low extraction accuracy.We proposed a convolu-tional neural network based on deformable convolution(D-Net),aiming to accurately extract road areas from remote sensing images.To validate the model’s performance,we conducted experiments on the road datasets of Massachusetts and DeepGlobe.Additionally,we chose a GF-2 satel-lite image to evaluate the model’s generalization ability.Experimental results demonstrate that D-Net significantly improves road segmentation accuracy,providing strong support for optimizing extraction results through deformable convolutions.The progress of this research holds signifi-cant theoretical and practical implications for further exploration in intelligent road extraction.
作者
高玉慧
朱武
张锐煊
王越
GAO Yuhui;ZHU Wu;ZHANG Ruixuan;WANG Yue(Yunnan Communications Investment&Construction Group Co.,Ltd.,Kunming 650103,China;School of Geological Engineering and Geomatics,Chang’an University,Xi’an 710054,China)
出处
《地理空间信息》
2024年第6期25-28,共4页
Geospatial Information
基金
国家重点研发计划项目(2020YFC1512000)
陕西省杰出青年基金项目(2023-JC-JQ-24)。
关键词
形变卷积
深度学习
遥感技术
道路提取
deformable convolution
deep learning
remote sensing technology
road extraction
