改进的U-Net卷积网络在遥感影像地物分类中的应用  

Improved U-Net convolutional network application for land cover classification in remote sensing images

作　　者：苟长龙[1] 庞敏 杨扬 GOU Changlong;PANG Min;YANG Yang(Gansu Vocational College of Communications,Lanzhou 730207,China;Shanxi Institute of Surveying and Mapping Geographic Information,Taiyuan 030001,China)

机构地区：[1]甘肃交通职业技术学院,甘肃兰州730207 [2]山西省测绘地理信息院,山西太原030001

出　　处：《测绘通报》2025年第3期150-155,共6页Bulletin of Surveying and Mapping

基　　金：2021年甘肃省高等学校创新基金(2021B-463)。

摘　　要：地物分类在环境监测、资源管理和城市规划中具有重要作用,但光谱相似性、噪声干扰及自然与人造地物混杂等因素,使得分类过程面临各种挑战。为提高分类精度,并增强模型的稳健性,本文提出了一种基于U-Net卷积网络架构且结合Transformer自注意力机制的深度学习网络。在兰州市遥感影像数据集上的试验表明,该模型在平均分类精度(mAcc)、平均交并比(mIoU)和平均F1分数(m F1)等指标上均优于PSPNet、DeeplabV3、Segformer和Swin-T模型。该模型不仅提高了分类精度,还实现了较高的推理速度,展现出在复杂地物场景中的应用潜力,为遥感影像分类提供了新思路。Land cover classification plays a crucial role in environmental monitoring,resource management,and urban planning.However,the classification process faces various challenges due to factors such as spectral similarity,noise interference,and the intermixing of natural and man-made objects.To improve classification accuracy and enhance the robustness of the model,this paper proposes a deep learning network based on the U-Net convolutional architecture,combined with the Transformer self-attention mechanism.Experiments conducted on the remote sensing dataset of Lanzhou city demonstrate that the proposed model outperforms PSPNet,DeeplabV3,Segformer,and Swin-T in terms of average classification accuracy(mAcc),mean intersection over union(mIoU),and mean F1 score(m F1).In addition to improving classification accuracy,the model achieves high inference speed,showcasing its potential for applications in complex land cover scenarios and offering new insights for remote sensing image classification.

关 键 词：深度学习 地物分类 卷积神经网络 遥感影像 语义分割 

分 类 号：P237[天文地球—摄影测量与遥感]