Establishment Of The LCZ Dataset For Large Cities In China For High-resolution Urban Climate Simulations

High-resolution urban building classification data is critical to the development of urban meteorology and related disciplines.The Local Climate Zone(LCZ)scheme is a new method to construct urban high-resolution land use classification data.Based on LCZ classification system,this paper firstly produced LCZ dataset of 63 major cities and urban agglomerations in China based on the currently widely used World Urban Database and Access Portal Tools(WUDAPT)method.In further study,in order to improve the efficiency and accuracy of LCZ map production,deep learning(DL)algorithms were used,including the improved Sen2LCZ-NET and Residual network(Res Net)based on convolutional neural network(CNN).With different algorithms,the LCZ dataset of Beijing and Guangzhou were produced,and the accuracy of LCZ maps based on different methods were discussed.The main conclusions are as follows:(1)This research presented a LCZ dataset for 63 cities and 4 major agglomerations in China.The dataset coupled the World Urban Database and Access Portal Tools(WUDAPT)project based on Landsat 8 images from 2019 with the LCZ classification method.76000training samples were used to provide spectral features and 23000 validation samples were used to ensure unbiased accuracy assessment.The results from the quality assessments showed that the LCZ dataset is generally of good quality,with an overall accuracy(OA)between 71%and93%among cities and regions,and an average OA of 82%.The accuracy of built types(OA_u)ranges from 57%to 83%,with an average of 72%.The accuracy of natural types(OA_n)ranges from 70%to 99%,with an average of 90%.(2)The results show that the quality and quantity of training samples are the key to the high accuracy of LCZ map in the WUDAPT method.The larger the number of homogeneous training samples,the higher the accuracy.When the number of training samples approaches saturation,the accuracy of LCZ map mainly depends on the urban form.In general,the more regular the urban form,the higher the accuracy of LCZ map.Among the all built types identified by WUDAPT,the accuracy of lightweight low-rise building(LCZ 7)and compact middle-rise building(LCZ 2)is higher.In addition,the small scale of some LCZ types,the similarity between LCZ types and the continuous distribution of LCZ types in the actual underlying surface are the reasons for the low accuracy of some built types.(3)Based on Sentinel-2 images from 2019 and the So2Sat LCZ42 training dataset,the deep learning models Sen2LCZ-NET and Res Net based on CNN were used to classify Beijing and Guangzhou.The LCZ classification results show that the LCZ underlying surfaces obtained by Sen2LCZ-NET and Res Net are generally consistent,but there are slight differences in some regions.Compared with the WUDAPT method,LCZ map obtained by the deep learning method is more coherent and the degree of fragmentation is small.(4)Among the three methods of WUDAPT,Sen2LCZ-NET and Res Net,Sen2LCZ-NET had the best classification performance.According to the accuracy assessment,the OA for LCZ maps in Beijing and Guangzhou based on SEN2LCZ-NET are 80%and 91%,respectively.Compared with WUDAPT,the OA of Beijing and Guangzhou have increased by 8%and 15%,respectively.Res Net was the next best,with an OA of 74%and 86%for LCZ maps in Beijing and Guangzhou,respectively.Compared with WUDAPT,the OA of Beijing and Guangzhou have increased by 2%and 11%,respectively.For classification accuracy,most LCZ types have higher accuracy in Sen2LCZ-NET,especially for compact LCZ types(LCZ 1-3).However,in the deep learning algorithm,there are still a some LCZ types that are confused due to the similarity of two-dimensional optical images in appearance.(5)Deep learning not only has a great improvement in the accuracy of maps,but also has great advantages in the efficiency of mapping.In terms of data source acquisition,a large scale of cloudless Sentinel-2 images can be aggregated based on Google Earth Engine(GEE),which provides a reliable data source for deep learning to produce LCZ maps.In terms of method,the combination of Sen2LCZ-NET and the existing So2Sat LCZ42 training dataset greatly improves the efficiency and accuracy of LCZ mapping,and provides an idea for constructing LCZ datasets for a larger area of high-resolution urban climate simulation in the future.