摘要
传统的土地利用分类方法大多基于对资料或影像的人工解译,存在一定的局限性。近年来,结合空间大数据和自然语言处理技术进行低成本快速的土地资源管理已成为研究热点。以美国纽约市曼哈顿区为例,提出了融合遥感影像和社会感知数据的城市土地利用分类方法。从遥感影像中提取光谱特征、从推特数据中提取用户活动时空和主题特征,基于随机森林法和深度神经网络法,构建了细粒度的城市土地利用分类模型。通过对比不同特征组合分类方法的精度,得到结合光谱特征和用户活动时空、主题特征的深度神经网络方法的结果最优,总体精度达82.65%,Kappa系数为70.1%。结果表明,社会感知数据中隐含的用户活动时空模式和活动主题信息均有助于提高城市土地利用分类的精度,而神经网络法可有效融合多源数据,为快速、低成本获取城市土地利用信息提供了新的途径。
Traditional land use classification methods are mostly based on labor-intensive interpretation of image,which have certain limitations.In recent years,integrating big data and natural language processing technology to carry out low-cost and rapid land resource management has become a hot issue.Take Manhattan as an example,this paper studies the urban land use classification based on remotely sensed and social sensed data.The spectral features of remotely sensed image,the spatiotemporal pattern of twitter user trajectory and the latent topics of tweet content related to user activity are extracted.Two common classification methods,random forest and deep neural network,are applied to construct urban land use classification models.The highest accuracy is obtained by deep neural network method based on remotely sensed and social sensed data,with overall accuracy at 82.65%,and Kappa at 70.1%.The results show that both spatiotemporal and textual features extracted from social sensed data are of great importance in urban land use classification.And deep neural network can integrate information from multi-source data,which provides a potential way for effectively classifying urban land use with open-source data.
作者
吴郁文
林杰
WU Yuwen;LIN Jie(School of Earth Science,Zhejiang University,Hangzhou 310027,China;Institute of Geography and Spatial Information,Hangzhou 310027,China)
出处
《浙江大学学报（理学版）》
CAS
CSCD
北大核心
2023年第1期83-95,107,共14页
Journal of Zhejiang University（Science Edition）
基金
国家自然科学基金资助项目(41501423)。
关键词
土地利用分类
遥感
社会感知
随机森林
深度神经网络
land use classification
remotely sensed
social sensed
random forest
deep neural network
