李夏1，李 媛2，张 博3

　　(1.水利部信息中心，北京 100053;

　　2.北京化工大学化学工程学院，北京 100029;

　　3.航天科工集团智慧产业发展有限公司，北京 100854)

　　摘要：生产建设项目地块数量多而分布分散，当前开展自动化遥感提取生产建设项目地块的研究较少，仍主要采用人工目视解译遥感影像的方法进行提取，存在效率低、成本高、稳定性差等问题。基于高分一号(GF-1)遥感影像，分析归纳不同类型和阶段生产建设项目组成地物的光谱、形状和空间特征，将防尘网确定为在建生产建设项目检测的特征地物，比较分析不同场景下DeepLabv3+和PSPNet 2种深度学习方法对防尘网的提取结果。研究结果表明：不同难易程度场景下，DeepLabv3+模型的识别结果均明显优于PSPNet模型，从而为在建生产建设项目防尘网的遥感监管提供一种新模式。快速获取生产建设项目地块的位置和边界信息，对提升水土保持监管效率、保护区域水土资源具有重要意义。

　　关键词：图像融合;卷积神经网络;DeepLabv3+;PSPNet;自动识别;适用性;防尘网

　　Feasibility study of DeepLabv3+ and PSPNet algorithms for automatic identification of dust screens

　　LI Xia1，LI Yuan2，ZHANG Bo3

　　(1. Information Center, Ministry of Water Resources, Beijing 100053, China;

　　2. College of Chemical Engineering, Beijing University of Chemical Technology, Beijing 100029, China;

　　3. Smart Industry Development of China Aerospace Science and Industry Co., Ltd.，Beijing 100854, China)

　　Abstract: The sites for production and construction projects are numerous and scattered. At present, there is few research on automatic remote sensing extraction of sites for production and construction projects. Manual visual interpretation of remote sensing images is still the main method for extraction, which has problems such as low efficiency, high cost, poor stability. Based on GF-1 satellite images, this paper analyzes and summarizes the spectral, shape and spatial features of different types and stages of production and construction sites. Dust screens are selected for identification of production and construction projects, and extraction results of the dust screens by DeepLabv3+ and PSPNet in different scenarios are compared and analyzed. The research results show that identification results of the DeepLabv3+ model are significantly better than that of the PSPNet model in all scenarios, which provides a new model for remote sensing supervision on dust screens of production and construction projects. It is of great significance to quickly obtain the location and boundary information of production and construction project sites for improving the supervision efficiency of soil and water conservation and protecting regional water and soil resources.

　　Key word：image fusion;convolution neural network;DeepLabv3+;PSPNet;automatic identification;feasibility;dust screens