摘要
准确掌握城市绿地的面积和空间分布对于城市园林绿化规划和管理具有重要意义。针对U-Net网络提取城市绿地存在的参数冗余和边界特征丢失等问题,本研究提出一种基于改进U-Net网络的城市绿地信息自动提取方法。该方法采用MobileNetv2作为编码部分,结合交叉熵损失函数和Dice损失函数以提高模型的泛化性并解决样本不平衡的问题。同时,引入空间通道压缩与激励模块以解决边界特征提取不准确的问题,在提高模型精度和速度的同时,降低参数量。结果表明:(1)改进的U-Net模型在城市绿地提取的精度和速度方面优于其他4种经典模型,参数量仅为6.9M;相较于原始的U-Net模型,改进后U-Net模型的mIoU提高0.63%,参数量减少77.77%,平均帧数提高2.77倍,表明该方法在显著减少模型参数的同时,仍能保持较高的精确性。(2)实际应用方面,研究部署最优模型对长沙市区的绿地进行自动提取,共提取75549个绿地图斑,面积范围从0.0001hm^(2)到706.39hm^(2)不等;分区统计结果显示,岳麓区绿地面积最大,芙蓉区最小,与目视解译结果基本一致。本研究不仅提供了一种改进城市绿地提取精度和速度的方法,而且具有一定的实用价值,为城市绿化规划和管理提供有力的数据支持。
Acquiring the layout and quantity of urban green space is of significant importance for the planning,construction,and management of urban gardens and greenery.The U-Net network applied to the extraction of urban green space encounters issues such as parameter redundancy and the loss of boundary features.To address these issues,this study proposed an improved automatic extraction method for urban green space information based on the U-Net model.This method employed MobileNetv2 as the encoder and combined the Cross-Entropy Loss Function with the Dice Loss Function to enhance model generalization and address sample imbalance.It introduced a spatial and channel squeeze-and-excitation module to solve the problem of inaccurate boundary feature extraction,thereby improving the model's accuracy and speed while reducing the number of parameters.The results demonstrated as follows:(1)The improved U-Net model outperformed four other classic models in terms of urban green space extraction accuracy and speed,with only 6.9 M parameters:Compared to the original U-Net model,the mloU increased by 0.63%,the number of parameters decreased by 77.77%,and the average frame rate increased by 2.77 times,indicating that the method significantly reduced the model's parameter count while ensuring accuracy.(2)In application,the deployment of the optimal model to extract green space in Changsha city identified 75549 green patches,with areas ranging from 0.0001 hm’to 706.39 hm^(2).Statistical analysis showed that Yuelu District had the largest green space area,while Furong District had the smallest,which was largely consistent with the results of visual interpretation.Therefore,the outcomes of this study not only offered an improved method for enhancing accuracy and speed in urban green space extraction tasks but also possessed practical utility and could provide data support for urban greening planning and management.
作者
张国珍
雷昌龙
严恩萍
杨明
刘丽娜
钟雅婷
ZHANG Cuozhen;LEI Changlong;YAN Enping;YANG Ming;LIU Li'na;ZHONG Yating(Hunan Academy of Building Research Co.,Ltd.,Changsha 410022,Hunan,China;Research Center of Forestry Remote Sensing&Information Engineering,Central South University of Forestry&Technology,Changsha 410004,Hunan,China;Hunan Provincial Key Laboratory of Forestry Remote Sensing Based Big Data&Ecological Security,Changsha 410004,Hunan,China;Key Laboratory of National Forestry and Grassland Administration on Forest Resources Management and Monitoring in Southern China,Changsha 410004,Hunan,China;Hunan Academy of Forestry,Changsha 410004,Hunan,China)
出处
《湖南林业科技》
2024年第3期10-18,共9页
Hunan Forestry Science & Technology
基金
湖南省林业科技创新资金项目(XLK202108-8)。
