Urban land cover mapping under the Local Climate Zone scheme using Sentinel-2 and PALSAR-2 data

Highlights

• Combining optical and SAR can improve urban land cover classification accuracy.

• We mapping urban land cover in Shanghai based on Local Climate Zone (LCZ) scheme.

• The subspace method is superior to SVM and MLC using high-dimensional data.

• Combining opticaland polarimetric features can achieve high classification accuracy using the subspace method.

Abstract

The spatial and spectral heterogeneity of urban areas makes land cover classification a challenging process. In this study, we highlight the potential of combined multi-spectral Sentinel-2 and fully polarimetric PALSAR-2 data for land cover classification in dense urban areas, based on the Local Climate Zone (LCZ) scheme. We classified differently combined spectral and back-scattering characteristics using the subspace method in comparison with the Support Vector Machine (SVM) and Maximum Likelihood Classifier (MLC) methods. Results show that, (i) the overall accuracy (OA) was 65.9% for the Sentinel-2 data, (ii) higher OA (71.9%) was achieved by adding four intensity images of PALSAR-2 to Sentinel-2, (iii) the inclusion of decomposed components increased OA to 72.8%, and (iv) the highest OA (73.3%) was achieved using all features. These results suggest that the inclusion of different backscattering characteristics disproportionately improved classification accuracy from using multi-spectral data alone. The results of comparison between different methods show that the subspace method performed better than SVM and MLC, particularly when high-dimensional data were used. The subspace method classified particularly well for some specific LCZ classes which are easily mixed between each other. It provides a promising option for LCZ mapping.

Introduction

With the increasing concentration of population in urban areas, the conversion from natural to urban land cover induces changes in the physical characteristics of the urban surface (e.g., albedo, thermal capacity, heat conductivity, and moisture content) (Voog and Oke, 2003). The urban heat island (UHI) effect, which is characterized by higher air and surface temperature in urban areas than in surrounding rural areas, is regarded as the most significant consequences of these changes (Voog and Oke, 2003). As it affects health and well-being of urban residents as well as urban environments, the UHI is a recent “hot topic” for research worldwide (Mohajerani et al., 2017).

Stewart and Oke (2012) were the first to develop a standardized worldwide land cover classification system to describe physical properties of urban morphology and corresponding thermal performance in UHIs, namely the climate-oriented Local Climate Zone (LCZ) scheme. As a standardized framework to characterize urban forms and functions, the LCZ scheme is applied globally as part of the World Urban Database and Portal Tools (WUDAPT) initiative (Danylo et al., 2016). The LCZ scheme has gone beyond its original purpose of studying the UHI phenomena and is now recognized as a valuable framework for characterizing urban land cover with satellite remote sensing images (Ching, 2013; Qiu et al., 2019).

Because of the spatial and spectral heterogeneity of urban areas, separating different land cover types with spectral characteristics using only optical images is a challenging process (Cockx et al., 2014). Recent studies demonstrated that synthetic aperture radar (SAR) data can provide complementary information to optical data (Darmawan et al., 2015), and their integration often leads to increased classification accuracy (Bechtel et al., 2016; Xu et al., 2017a; Xu et al., 2017b; Pavanelli et al., 2018).

Fully polarimetric SAR collects complete backscattering information in the coherency matrix T₃. Based on T₃, different target decomposition methods were developed to transform the backscattering information into basic backscattering mechanisms for geophysical parameter inversion as well as terrain and target classification (Cloude and Pottier, 1997; Touzi et al., 2004; Yamaguchi et al., 2005). Therefore, fully polarimetric SAR offers a powerful way to detect and characterize targets, based on electromagnetic radiation characteristics and backscattering mechanisms (Lee and Pottier, 2009; Muneyoshi and Junichi, 2013; Angelliaume et al., 2018; Ohki and Shimada, 2018).

Recent studies, taking advantage of the spectral, spatial, and polarimetric information, have combined optical and fully polarimetric SAR data to improve land cover classification accuracy (Bagan et al., 2012; Zhu et al., 2012; Merchant et al., 2019; Muthukumarasamy et al., 2019). However, the integration of backscattering intensity, decomposed components and polarimetric parameters from fully polarimetric SAR and multi-spectral imagery for land cover classification needs further investigation.

A dataset combined by multi-source data is likely to be imbalanced and with data redundancy when the dimensionality is relatively high. Using such a dataset for land cover classification would be a challenge. The subspace method has shown potential for classifying remote sensing data with high-dimensionality, due to the ability of reducing data dimensionality and extracting the features simultaneously (Bagan et al., 2012). To demonstrate the proposed subspace method, we applied it to datasets integrated by Sentinel-2 and fully polarimetric PALSAR-2 data in Shanghai, China, and investigated the robustness of the subspace method in comparison with the Support Vector Machines (SVM) (Mountrakis et al., 2011; Koda et al., 2018) and Maximum Likelihood Classifier (MLC) (Lillesand et al., 2015).

The remainder of this paper is structured as follows. In Section 2, we described the study area, Seninel-2 and PALSAR-2 imagery, ancillary data, LCZ scheme, classification method, classification process and accuracy assessment. Section 3 shows the experimental results. In Section 4, we analyzed the experimental results and discussed the remaining challenges and possible solutions for future work. Finally, Section 5 summarizes and concludes our study.