Urban Land Use Classification Based On A Few Labeled Samples

With the rapid development of economy and strong policy support,the current process of urbanization is accelerating.The classification of urban land use plays an important role in urban planning and construction.It can be used to guide the allocation of urban resources and understand the distribution of urban regions,so as to help the construction of smart city.In the study of urban land use classification,the collection and marking of samples require field investigation and visual interpretation,which leads to the time and labor consuming for collecting a large number of marked samples and the easy occurrence of errors.Therefore,this paper mainly focuses on the classification of urban land use in Shenzhen under the condition of few labeled samples.The main research work of this paper includes the following contents:1.The classification of urban land use in the study area is studied by using multisource geographic big data.Sentinel-2 high resolution remote sensing image data,Luojia1-01 night-time light remote sensing image data,mobile phone location data and Gaode POI data are collected.In this paper,data cleaning,coordinate conversion,projection and format conversion,data calculation and integration,statistics and other processing are carried out on the above data.The features of different data are integrated and processed on each land parcel,and the features are analyzed,processed and selected to complete the analysis and processing of data and samples.2.Select and build several classification models and compare them,and improve the Co-Forest semi-supervised classification method according to the scene.Compared with traditional land use classification methods such as field investigation and remote sensing image interpretation,machine learning is selected in this paper to classify land use,including model construction of supervised classification(Random Forest,XGBoost)and semi-supervised classification(Co-Forest)methods,and further improved the Co-Forest classification method from many aspects.3.The determination of the proportion of stable training samples corresponding to urban land use classification in the research area under the condition of a few labeled samples.This paper designs experiments within the scope of the training sample of 12% to 1%,uses three kinds of classification methods to research urban land use classification in study area,compares the results of different classification methods,and determines the stable proportion of the training sample,in addition,further research on multi-source geographic data for urban land use classification in the degree of importance.4.The practice and application of urban land use classification.The model constructed under the condition of few labeled samples is put into practice in Shenzhen,that is,the 5% training sample proportion is used to classify the land use in Shenzhen by XGBoost and the improved Co-Forest classification methods,and the practical application direction is further elaborated and expanded.Through experiments and analysis in this paper,according to the overall classification results of the study area,the improved semi-supervised classification method Co-Forest has the best classification result,and XGBoost is similar to it,while the Random Forest algorithm most used by researchers has a slightly worse performance.The experimental results show that the Co-Forest algorithm is improved greatly.In terms of the stability of the sample proportion,the experiment in this paper shows that the training sample proportion of 5% or above can obtain a relatively stable classification effect in the study area.The importance of single geographic data features is as following: POI data> Sentinel-2 high-resolution remote sensing data >mobile phone data > NTL remote sensing data.The classification effect of using multi-source geographic data is better than single-source geographic data.The practice and application direction of land use classification in Shenzhen also further shows the adaptability of the research in reality,so as to help urban planning and construction.