Urban Vegetation Classification And Biomass Inversion Based On Sentinel-2A Data In Xuzhou

Research on urban vegetation classification and biomass inversion is not only the basis for urban landscape planning and management,but also provides data source and technical support for the studies on urban vegetation ecological regulation and environmental protection.Using Sentinel-2A remote sensing imagery and field sites survey data,this thesis explores urban vegetation classification methods and biomass inversion models with the vegetation within the Sanhuan Road of Xuzhou.Through variable selection,four different combinations of classification variables were determined according to the types of variables involved in the classification.Then those four combinations were applied to five classifiers for image classification.In terms of urban vegetation biomass inversion,three types of biomass inversion models,namely the simple regression model?SR?,the multiple linear regression model?MLR?,and the stepwise regression model?Both-SRA?,were established for each vegetation type with the field-based biomass survey data.According to the optimal biomass inversion model identified,biomass distribution and annual variation of the study area were analyzed with six-temporal Sentinel-2 satellite imager data in 2017.The main research results and conclusions are as follows:?1?Urban vegetation endmember information was extracted.There were three steps before performing a liner spectral mixture analysis?LSMA?.The first step was minimum noise fractionation?MNF?,followed by vegetation endmembers“primary screening”and candidate endmembers“purification”.After LSMA,three vegetation endmembers,namely low vegetation,broad-leaved forest,and coniferous forest were obtained.The abundance maps of the three vegetation components were produced by the fully constrained least squares?FCLS?method,and the root mean square?RMS?was 0.019 which means the LSMA result was acceptable.?2?Urban vegetation classification methods were investigated.With the same classification variables,the support vector machine?SVM?classifier resulted in the highest classification accuracy in the five classifiers,the random forest?RF?,artificial neural network?ANN?,and quick unbiased efficient statistical tree?QUEST?classifiers got the second classification accuracy,and the MLC was ranked in the last.For the same classifier,we found that the classification accuracy can be improved with the addition of variables?i.e.,from M1 to M4?.Among all the classification models based on five classifiers,the SVM_M4 model got the best accuracy with an overall accuracy of 89.86%and a Kappa coefficient of 0.83.The classification accuracy of low vegetation,broad-leaved forest,and coniferous forest were 86.76%,90.41%,and 91.43%respectively.?3?Urban vegetation biomass inversion models were established.Among the three modeling methods,the Both-SRA model resulted in the highest fitting and prediction accuracy.Model fittings(R²_nh)of low-vegetation,broad-leaved forest and coniferous forest were 0.853,0.821,and 0.838 respectively.Model determination coefficient(R²_yz)were 0.7679,0.7318,and 0.7860 respectively.In addition,results also show that,in terms of fitness and prediction accuracy,models obtained by distinguishing the vegetation types were superior to the models without.?4?The biomass changes of urban vegetation in Xuzhou were analyzed.Different vegetation biomass can be inferred by using the Both-SRA model.Results show that the total biomass of the study area in satellite imagery acquisition time?July 2017?was2.18×10⁵t.Among them,the coniferous forest was the highest,which was 1.30×10⁵t and followed by the broad-leaved forest?6.20×10⁴t?.Low vegetation biomass was lowest ranked with 2.64×10⁴t.Annual biomass analysis show that the study area had the lowest biomass in January and December,and the highest biomass in September.