Combing Optical Image And Full-polarimetric SAR Data For Urban Land Cover Classification

The spatial and spectral heterogeneity of urban areas makes classification of land cover types a challenging process.It is hard to classify land cover types with similar spectral characteristics using multi-spectral remote sensing images solely.Full-polarimetric SAR records complete scattering information of targets on Earth's surface,including intensity,structure and physical properties.Taking advantage of complementarity between multi-spectral images and full-polarimetric SAR data,we could combine multi-spectral information,intensity of scattering information in different polarization,features extracted from different target polarimetric decomposition methods and complete polarimetric parameters of the coherency matrix T3 into a dataset.Using the combined dataset to classify land cover types in urban area with high complexity,the categories with high spectral similarity could be separated well.This study highlights the joint use of multi-spectral Sentinel-2 MSI imagery and full-polarimetric ALOS-2 PALSAR-2 data to map 16 land cover categories,based on the Local Climate Zone(LCZ)scheme,in Shanghai,China.Firstly,we combined spectral bands,scattering intensity data,target polarimetric decomposed components,and scattering parameters of the coherency matrix T3 into six datasets for land cover classification.Secondly,we investigated land cover discrimination of different datasets using the subspace classification method,compared to the Support Vector Machine(SVM)and Maximum Likelihood Classifier(MLC)methods.We then calculated the confusion matrix of land cover classification maps categorized by 16 types.Also,we merged built-up types into a single type and then obtained corresponding classification maps with 9 classes.The confusion matrices of merged maps were calculated.Finally,overall accuracy(OA),producers' accuracy(PA)and users' accuracy(UA)were derived from confusion matrix as accuracy measurements.By comparing and analyzing these three measurements,we could investigate discrimination capacity of different land cover types for six datasets.The comparison of classification performance of subspace method,SVM,and MLC shows that SVM performed best when using datasets composed of 12 spectral bands.In contrast,the subspace method performed better than SVM or MLC using datasets with high dimensionality or with data redundancy.Comparing the 16 land cover types classification overall accuracy(OA)using the subspace method shows that,1)with the Sentinel-2 data,the OA was only 65.9%,2)higher OA(71.9%)was achieved by adding four intensity images of ALOS-2 PALSAR-2 to Sentinel-2,3)the inclusion of decomposed components increased OA to 72.8%,and 4)the highest OA(73.3%)was achieved using all multi-spectral and full-polarimetric SAR features.The highest OA of classification map having 9 classes was 83.3%,generated from dataset with highest dimensions by subspace method.Notably,the corresponding PA of built-up class was up to 94.1%.The classified results of datasets by subspace method suggests that full-polarimetric imagery improves the discrimination capability of land cover classes,compared to using multi-spectral data solely.