Multi-temporal Sentinel-1/2 And Google Cloud Service In Support Of Object-based Land Cover Mapping

Accurate land cover mapping and change monitoring in high spatial and temporal resolution is of great significance to land use planning and ecological environment monitoring.The promise of the combined use of optical and microwave data for improving classification performance has been fully investigated in the related literature: it can significantly improve the accuracy of land cover mapping relative to a single data source.However,the past researches tend to focuses on the single-temporal,pixel-oriented classification,and the investigation of the multi-temporal,object-oriented classification,especially in urban scenario,is not sufficient.In this study,utilizing the data processing capability and ready-to-use Sentinel catalog provided by Google Earth Engine,nearly 600 features in spectral,spatial,temporal,geometric and polarimetric dimensions were generated.Then,they were reduced to 63 most discriminate features using random forestbased recursive feature elimination method,as inputs to 3 popular machine learning algorithms in remote sensing community-Support Vector Machine,random forest and KNearest Neighbors.Then,ensemble analysis was conducted of the outputs from the three classifiers.Its overall accuracy was 0.816 and kappa coefficient was 0.786,which was significantly superior than SVM,the best classifiers among the three in terms of overall accuracy and kappa coefficient(overall accuracy of 0.738 and kappa coefficient of 0.695).In comparison experiment,multi-temporal synthetic use of optical and SAR data with feature selection,our proposed approach,achieved an overall accuracy of 0.720,while the single-temporal optical data 0.671,multi-temporal optical data 0.681,and multi-temporal synthetic use of optical and SAR data without feature selection 0.698.The guidance of the uncertainty map that comes along with ensembled land cover map,on indicating the spatial distribution of error was validated in accuracy assessment using stratified sampling,in which accuracies of the high,medium and low confidence level are 100%,74%,44%.