BCI: A biophysical composition index for remote sensing of urban environments

Abstract

Understanding urban environments and their spatio-temporal changes is essential for regional and local planning and environmental management. To facilitate monitoring and analyzing urban environments, remotely sensed data have been applied extensively because of its synoptic view and repeat coverage over large geographic areas. Compared with traditional per-pixel and sub-pixel image analyses, spectral indices have apparent advantages due to their easy implementation. However, most spectral indices are designed to highlight only one land cover, and confusion between other land cover types, in particular impervious surfaces and bare soil, has not been successfully addressed. This study proposes a biophysical composition index (BCI) for simple and convenient derivation of urban biophysical compositions for practical applications following Ridd's conceptual vegetation – impervious surface – soil triangle model by a reexamination of the Tasseled Cap (TC) transformation. Further, this research explores the applicability of BCI in various remotely sensed images at different spatial resolutions. Results indicate that, BCI has a closer relationship with impervious surface abundance than those of normalized difference vegetation index (NDVI), normalized difference built-up index (NDBI) and normalized difference impervious surface index (NDISI), with correlation coefficients of approximately 0.8 at various resolutions. Also, the performances of BCI in quantifying vegetation abundance are comparable with NDVI at all three spatial scales. Additionally, with much higher values of separability metrics than any other index, the study confirms that BCI was shown to be the most effective index of the four evaluated for separating impervious surfaces and bare soil.

Introduction

Over the past decades, urbanization has taken place at an unprecedented rate around the world. In 1950, as an example, only 29% of the world population resided in urban areas, and this number rose to 49% by 2005. Moreover, this trend is projected to continue in the future decades (United Nations, 2006). In the process of urbanization, natural landscapes have been rapidly transformed to anthropogenic urban land uses, and biophysical compositions and characteristics of natural environments have been dramatically modified. While urbanization brings social and economic benefits (e.g. improved quality of life and economic prosperity), it also leads to a number of environmental problems (e.g. water quality degradation, air pollution, loss of biodiversity, and urban heat island effect) and social issues (e.g. excess commuting, economic and social inequality, congestion, etc.) (Xian & Crane, 2006). Due to these significant impacts, understanding urban environments and their spatio-temporal changes is essential for regional and local planning and environmental management.

Remote sensing techniques provide an important means for understanding urban environments. With a synoptic view and repeat coverage of a large geographic area, remotely sensed data have been applied extensively to analyze urban environments. Traditionally, remote sensing imagery has been employed to derive multi-temporal land use land cover (LULC) maps with numerous spectral, spatial, and contextural analytical algorithms (Yuan & Bauer, 2007). In particular, the United States Geological Survey (USGS) has developed multi-temporal national land use land cover (NLCD) datasets with the help of Landsat Thematic Mapper (TM)/Enhanced Thematic Mapper Plus (ETM +) imagery (Yang et al., 2001). In addition to traditional land use land cover classification approaches, sub-pixel analysis has been developed following the vegetation – impervious surface – soil (V–I–S) model proposed by Ridd (1995). According to this conceptual framework, all land cover types (other than water) in an urban environment can be regarded as a combination of three basic biophysical components, namely vegetation, impervious surfaces, and soil. On this basis, two major categories of methods were developed to quantify biophysical compositions in an urban area. The first category is machine learning methods, including artificial neural network (ANN) (Flanagan and Civco, 2001, Hu and Weng, 2009, Mohapatra and Wu, 2007, Pu et al., 2008), regression/decision tree method (Lu and Weng, 2009, Mohapatra and Wu, 2010, Xian and Crane, 2005, Yang, Huang, Homer, Wylie and Coan, 2003, Yang, Xiang, Klaver and Deal, 2003, Yuan et al., 2008), and regression modeling (Mohapatra and Wu, 2010, Wu and Yuan, 2007, Yang, 2006, Yang and Liu, 2005). With these machine learning methods, biophysical composition information is derived by establishing an empirical relationship with various spectral and spatial characteristics extracted from remote sensing imagery. The second category is spectral unmixing techniques (Lu and Weng, 2004, Phinn et al., 2002, Powell et al., 2007, Powell et al., 2008, Rashed et al., 2003, Roberts et al., 1998, Small, 2001, Small, 2005, Weng, 2012, Wu, 2004, Wu, 2009, Wu and Murray, 2003). Under the assumption that the spectrum of a pixel is a combination of the spectra of several typical homogeneous ground components, named endmembers (Adams et al., 1995, Roberts et al., 1998), the areal fractional coverage of each ground component can be derived from the spectral mixture analysis (SMA) approach.

Although both per-pixel and sub-pixel analyses have been employed for analyzing urban environments with different degrees of success, these methods are always considered as complicated, computationally intensive, and sometimes subjective, especially when applied to a large geographic area (Plaza et al., 2004, Somers et al., 2011). Taking the sub-pixel analysis as an example, the success of machine learning algorithms relies heavily on the quality of training and testing data, and the selection of which might be relatively subjective during the process of model building and validation (Yang et al., 2003a). Further, although the SMA algorithms are physically based approaches and able to acquire sub-pixel endmember fractions effectively (Franke et al., 2009, Powell et al., 2007, Roberts et al., 2012), it is extremely difficult for practical users because of its complicated implementation process. Such difficulties lie mainly in technical problems including endmember selection and intra-class variability quantification, etc. (Somers et al., 2011, Wu, 2009). One exception is the multiple endmember spectral mixture analysis (MESMA) method developed by Roberts and his colleagues (Franke et al., 2009, Powell et al., 2007, Roberts et al., 2012), although it is also considered to be computationally intensive. When compared with per-pixel and sub-pixel image analyses, spectral indices have apparent advantages due to their easy implementation and convenience in practical applications. A number of indices, including normalized difference vegetation index (NDVI; Rouse et al., 1974), soil adjusted vegetation index (SAVI; Huete, 1988), normalized difference built-up index (NDBI; Zha et al., 2003), and normalized difference impervious surface index (NDISI; Xu, 2010), have been developed to quantify biophysical characteristics of the earth's surfaces (Jackson & Huete, 1991). Although these indices have proven effective to some degree, when applied to urban environments, several problems still exist. The first problem is that most spectral indices are designed to highlight only one land cover type (e.g. vegetation and built-up area, etc.), and confusions among other land cover types, in particular impervious surfaces and bare soil, have not been successfully addressed. For instance, NDVI is designed for the signal enhancement of vegetation abundance, with which, however, impervious surfaces are always confused with bare soil. Moreover, although NDBI is intended to highlight built-up area, it cannot effectively differentiate built-up materials from barren soil (He et al., 2010). The second problem of spectral indices is associated with the limited applicability in remote sensing imagery at different spatial and spectral resolutions. Other than vegetation indices (i.e. NDVI and SAVI), all the aforementioned indices are unavailable for most high spatial resolution remote sensing imagery due to their dependence on shortwave infrared (SWIR) bands, which are not always included in high spatial resolution remotely sensed data (one exception is WorldView-2 imagery). Therefore, the objectives of this paper consist of (1) proposing a biophysical composition index (BCI) for simple and convenient derivation of urban biophysical compositions for practical applications by enhancing contrast and separability among different biophysical compositions, and (2) exploring the potential of the applicability of BCI in various remotely sensed imagery at different spatial resolutions.

The remainder of this article is organized as follows. The next section introduces the study areas and data. The third section presents the methodology of the BCI development, including a reexamination of Tasseled Cap components and their representations in urban environment, the formulation of BCI index, and the assessment of BCI's ability of representing urban biophysical compositions. Results of applying the BCI with Landsat ETM+, IKONOS, and MODIS imagery are reported in Section 4, and comparative studies with other three indices, namely NDVI, NDBI, and NDISI, are detailed in Section 5. Further, discussions are provided in Section 6, and finally, this paper concludes in Section 7.