Study Of Urban Green Space Extraction Based On Sentinel-1A And Sentinel-2A Data

As an important part of regulating urban ecological environment,landscape pattern,temperature,urban green space is of great significance to study its distribution in cities,at the same time,with the development of remote sensing technology,combining remote sensing data and radar data to extract features has great advantages.Some domestic and foreign scholars combined Sentinel-1A and Sentinel-2A to carry out land cover and other studies,but no research was conducted on urban green space.Therefore,this article aims to study the application of Sentinel-1A and Sentinel-2A image data in the extraction of urban green space.In this paper,the inner city of the Second Ring Road of Jinhua City is used as the research area,and Sentinel-1A and Sentinel-2A are used as data sources.Use Objectbased classification to classify land and focus on extracting green space,through six data schemes,explore the advantages of S2 red edge band,S1 radar data and texture features which compared to single data extraction in green space.At the same time,it compares the advantages of constructing class-level classification methods and directly using Objectbased classification methods,explores the optimal classification method for extracting green space from sentinel data,and analyzes the spatial distribution and seasonal variation of green space in the downtown area of the second ring road in Jinhua City.The main conclusions are:(1)S2A red edge band and S1 A radar data are both beneficial for urban land extraction,red edge band is particularly advantageous for green space extraction,and radar data is beneficial for urban artificial construction land extraction.Combining Sentinel-1A and Sentinel-2A data,using Object-based classification method,the classification algorithm selects Random Forest algorithm,which is divided into 7 types of land: construction land,transportation land,water,bare land,agricultural land,grassland and woodland,use six classify schemes.Scheme 1 only uses the S2 A seven visible light bands,Scheme 2 adds 3 red edge bands,Scheme 3 adds S1 A VV and VH bands,Comparing the first three schemes,the results show that the classification accuracy of scheme 1 is 78.99%,the Kappa coefficient is 0.75,and the classification result is not satisfied.The scheme 2 adds three S2 A red edge bands,and the classification accuracy reaches 80.93%,and the Kappa coefficient is 0.77,all have improved,and analysis the classification map,the extraction result of green space is obviously better;Scheme 3 adds VV and VH radar bands,with an accuracy of 84.44% and a Kappa coefficient of 0.81.The classification results are better,especially for the extraction of urban artificial construction land.(2)Adding features such as vegetation index texture is beneficial to ground feature extraction,but because all the features are involved in the classification,resulting in data redundancy,the classification effect is not obvious,and the classification results are better after using random forest for feature optimizationScheme 4 is the data of scheme 3 plus the red edge index,vegetation index,and S1A's S2A's texture features,42 features in total,the overall classification accuracy is improved by about 2% compared with scheme three,but there are still a lot of confusion on grassland agricultural land,Due to too many features,some features have a small contribution to the extraction of ground features,and the data redundancy caused by the addition will affect the classification effect.Therefore,scheme 5 first uses random forests to score feature importance,analyze the importance of various features,select the optimal feature combination,and then perform random forest classification.The classification results are the best.The classification accuracy reach 87.55%,and the Kappa coefficient is 0.85.However,the small green space between the houses and the greenhouse planting area are still not correctly extracted,so I add the six scheme which use layer classification method.(3)Analyzing the random forest importance evaluation results,the overall importance ranking of various features is: spectral feature> vegetation index> red edge index> S1 texture feature> S2 texture feature.Scheme 5 uses random forest to score the importance of features,and the results show:Among the top 10 features of importance score,there are 6 spectral features,and all spectral features are within the top 20,it proves the importance of spectral features;The vegetation index and the red edge index also rank high,contributing a lot to the classification;The overall ranking of texture features is lower than 20,but the texture features of S1 A radar are ranked higher than that of S2 texture features.It is proved that the texture features acquired by Sentinel radar data are superior to the texture features of Sentinel-2.In the end,the optimal feature combination for classification is the top 25 features,all the spectrum,vegetation index,and red edge index are selected,and only 3 S1 A texture features are selected.(4)Among the six schemes,scheme 6 which combined the object-based classification method with class hierarchy,the classification results are optimal,especially in the extraction of green spaces between buildings.Due to the complexity of urban features,different features are suitable for segmentation at different scales.Using a segmentation scale,there will be under-segmentation or oversegmentation of objects.Therefore,first construct four levels according to the division scale of 52,33,24,and 21,extract different land types at different levels,and classify them in order from large scale to small scale.Extract the water in L4;in L3,extract the construction land with less regular features such as industrial land,subdivides the green space in L2,and subdivides the central area with more mixed features in L1.Subdivide the central area with more mixed features in the L1 layer,and inheritance summary of all subcategories in the previous layers.The features in the L4 and L3 layers are easier to find the classification rules,so the spectral features of the features are directly analyzed,and the membership function is used for the rules classification;It is difficult to find rules directly for L2 and L1 subdivision,so use the five methods to classifier.In the end,the classification accuracy of the scheme reached 89.1%,and the Kappa coefficient reached 0.869.Comparing the classification results,this method makes the classification of ground features more accurate,especially for the extraction of ground features in small and complex areas.It is suitable for green spaces between buildings and agricultural land in greenhouses.The extraction is better,so scheme 6 is the optimal classification scheme.(5)Analyzing the spatial distribution and seasonal changes of green space in Jinhua City,green space accounts for 33.72% of the total land area,urban green space accounts for a relatively high area,grassland and agricultural land have significant seasonal changes,and woodland has little change.Use scheme six to classify the four-phase images of the four seasons in the study area,obtain a four-season classification map,and analyze the distribution and seasonal changes of green areas in Jinhua City.The research shows that the green area of Jinhua urban area accounts for 33.72% of the total area,and the urban green area accounts for a relatively high area.Among them,grassland is the most,agricultural land is second,and the woodland is the least.The grassland distribution is more scattered,agricultural land is distributed around the Second Ring Road,concentrated in the northwest and southeast of the urban area,and the woodland is concentrated in several larger parks.In the seasonal change of green space,grassland includes roadside green space,street trees,lawns in the park,and wild grassland.The overall seasonal change is large,with the most grassland in summer,followed by spring and autumn,and the smallest grassland area in winter,because most of the flaky grassland becomes barren land in winter;Agricultural land includes greenhouse cultivation and openair planting areas.The seasonal changes of the two are different,but the overall area in spring and summer is larger,followed by autumn and winter;The woodland is mainly formed by trees in the park,most of them are evergreen vegetation with relatively small seasonal changes.