Downscaling Land Surface Temperature in Complex Regions by Using Multiple Scale Factors with Adaptive Thresholds

Yingbao Yang, Xiaolong Li, Xin Pan, Yong Zhang, Chen Cao
2017 Sensors  
Many downscaling algorithms have been proposed to address the issue of coarse-resolution land surface temperature (LST) derived from available satellite-borne sensors. However, few studies have focused on improving LST downscaling in urban areas with several mixed surface types. In this study, LST was downscaled by a multiple linear regression model between LST and multiple scale factors in mixed areas with three or four surface types. The correlation coefficients (CCs) between LST and the
more » ... factors were used to assess the importance of the scale factors within a moving window. CC thresholds determined which factors participated in the fitting of the regression equation. The proposed downscaling approach, which involves an adaptive selection of the scale factors, was evaluated using the LST derived from four Landsat 8 thermal imageries of Nanjing City in different seasons. Results of the visual and quantitative analyses show that the proposed approach achieves relatively satisfactory downscaling results on 11 August, with coefficient of determination and root-mean-square error of 0.87 and 1.13 • C, respectively. Relative to other approaches, our approach shows the similar accuracy and the availability in all seasons. The best (worst) availability occurred in the region of vegetation (water). Thus, the approach is an efficient and reliable LST downscaling method. Future tasks include reliable LST downscaling in challenging regions and the application of our model in middle and low spatial resolutions. Also known as TIRS image sharpening or scale decomposition [10], LST downscaling used to be applied at the digital number (DN) level [11] or in the fusion of images [12] before 1994. Both the DN and surface temperature/radiance (Ts/Rn) levels existed simultaneously between 1995 and 2004, and Ts/Rn has prevailed since 2005 [13,14]. Most LST downscaling methods applied at the Ts/Rn level use either statistical regressions [10-18] or the modulation-based technique [10]. Statistical regressions connect LST with scale factors that are extracted from high-resolution, visible, near-infrared, or short-wavelength infrared bands through statistical correlations, which are often used because of their ease of use and acceptable downscaling accuracy. The most common statistics-based downscaling algorithms include the DisTrad method [15], the TsHARP algorithm [16-18], the PBIM algorithm [19], the EM algorithm [20], and the HUTS algorithm [21]. Statistical regressions can be either linear or non-linear, and linear regression formulas are known for their easy implementation and application. Tom et al. [22] pioneered the use of linear regression for downscaling LST, and their findings were further improved in Nishii [23]. However, linear regression formulas may not represent the nonlinear relationships between LST and scale factors. Therefore, piecewise linear and nonlinear regression models [24,25], conditional expectation models [23], co-Kriging models [26,27], and Bayesian models [28] have been established to represent the linear or nonlinear relationships between LST and one or more scale factors. When numerous scale factors exist, the increasingly complex relationships between LST and the scale factors can be effectively determined using artificial neural networks [29], genetic algorithm techniques [30], and support vector machines [31] that can incorporate multiple scale factors and identify hidden statistical patterns. However, these techniques have a limited generalizability and cannot discern a localized and clear physical relationship. Modulation-based downscaling methods establish a function of LST or thermal radiation brightness and land cover types based on the principles of thermal radiation and spectral mixture analyses [32,33]. Zhan et al. [34] proved that these methods could achieve an excellent downscaling effect. Most of the literature on LST downscaling have focused on the application of this technique at the Ts/Rn level. Different scale factors have been applied at varying application areas [35] and selected based on the characteristics of the study area. When the area is partly covered with vegetation, several vegetation indices can be used to downscale LST effectively [17,18], including the normalized difference vegetation index (NDVI) [15], fractal vegetation index [16-18], vegetation dryness index [2,36,37], and soil-adjusted vegetation index [38]. For instance, the widely adopted TsHARP method [17] uses NDVI in a linear regression model to downscale LST. However, these indices are unsuitable for LST downscaling in urban areas [19] [20] [21] 39, 40] . Given that each land cover type in cities has a unique emission rate, emissivity [19, 20] has been used as a major scale factor for highly heterogeneous urban areas. Small [41] found a close relationship between surface temperature and surface albedo in urban areas. Dominguez et al. [21] integrated NDVI and surface albedo to develop the HUTS algorithm. Impervious surface percentage [40, 42] and pure pixel index [43] have also been used as scale factors in urban areas. Essa et al. [40] found that substituting NDVI with impervious percentage in the DisTrad algorithm would produce better downscaling results by comparing 15 different scale factors. Essa et al. [40] and Yuan and Bauer [44] also found a strong linear relationship between LST and impervious percentage regardless of the season of image acquisition. By contrast, the relationship between LST and VI changes along with the seasons [15] . For complex urban areas with varying land cover types, multiple scale factors must be integrated to achieve a high downscaling precision. Although numerous approaches for LST downscaling have been proposed in the literature, they are often constrained by remote-sensing data and surface type. Moreover, an effective selection of scale factors for those areas with three or more land cover types is yet to be achieved. Our study proposes a multi-scale-factor downscaling method based on an adaptive threshold for those areas with complex land cover types. A detailed analysis of errors with spatial autocorrelation between the original LST image and the downscaled products is presented. The downscaled images are compared with the images obtained by other downscaling methods through visual and quantitative analyses. As its major contributions, this study uses multiple scale factors to downscale LST in
doi:10.3390/s17040744 pmid:28368301 pmcid:PMC5421704 fatcat:3u5in4545ng4rmg4mukiyc3nly