融合道路特征的动态卷积遥感影像道路提取网络模型  

A dynamic convolutional remote sensing image road extraction network model integrating road features

作　　者：陈涛 葛小三[1] CHEN Tao;GE Xiaosan(School of Surveying and Land Information Engineering,Henan Polytechnic University,Jiaozuo 454003,China)

机构地区：[1]河南理工大学测绘与国土信息工程学院,河南焦作454003

出　　处：《地理与地理信息科学》2025年第2期23-30,共8页Geography and Geo-Information Science

基　　金：国家自然科学基金项目(41572341);河南省自然科学基金项目(222300420450);河南理工大学研究生教育教学改革项目(2024YJ02)。

摘　　要：受道路特性和卷积池化等操作影响,现有遥感影像道路提取方法仍存在空间特征和地物细节信息丢失问题,造成提取错漏。该文基于遥感影像中道路狭长特点设计了一种语义分割网络模型OSPNet,基于条状注意力机制使模型更专注于道路信息,减少错提取;同时,使用动态卷积构造模型的编码器,从而增强模型对不同类型道路提取的泛化能力和对不同场景的适应性;在模型训练过程中,采用骰子损失函数和焦点损失的混合损失函数解决遥感影像中道路类别与背景类别不均衡的问题。在Massachusetts和CHN6-CUG道路数据集上的验证结果表明,与DeepLabV3+相比,该模型的F 1分数、交并比和平均交并比分别提升1.37%、1.78%和1.03%,相较其他网络评价指标结果最佳,提取的道路连续性、完整性好,且训练时间远低于其他模型,是一种适应性更强、分割精度更高、更轻量化的道路提取算法。Due to the influence of road characteristics and operations such as convolution and pooling,the existing methods still have problems of loss of spatial features and detailed ground object information,resulting in missed extraction and erroneous extraction.In this paper,a semantic segmentation network model OSPNet is designed for road extraction.Based on the characteristics of long and narrow roads in remote sensing images,this model uses a strip attention mechanism to make the model more focused on road information extraction and reduce erroneous extraction.At the same time,based on the diversity of roads in remote sensing images,dynamic convolution is used to construct the encoder of the model,thereby enhancing the generalization ability of the model for the extraction of different types of roads,and improving the adaptability of the model to different scenarios.During the model training process,a mixed loss function combining the dice loss function and the focal loss function is employed to address the issue of class imbalance between road and background categories in remote sensing images.In order to validate the road extraction ability of the model,the validation results on the Massachusetts road dataset and the CHN6-CUG road dataset indicate that compared with DeepLabV3+,the F 1-Score,intersection over union(IoU),and mean intersection over union(mIoU)of this model are increased by at least 1.37%,1.78%,and 1.03%respectively.The roads extracted by using this model have good continuity and integrity,and the training time is much shorter than that of other models.It is a more adaptable road extraction algorithm with higher segmentation accuracy and lighter weight.

关 键 词：道路提取 动态卷积 条状注意力机制 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP391.41[自动化与计算机技术—控制科学与工程]