Estimation of subpixel land surface temperature using an endmember index based technique: A case examination on ASTER and MODIS temperature products over a heterogeneous area

https://doi.org/10.1016/j.rse.2011.01.004Get rights and content

Abstract

Land surface temperature (LST) is a key parameter in numerous environmental studies. Surface heterogeneity induces uncertainty in estimating subpixel temperature. To take an advantage of simultaneous, multi-resolution observations at coincident nadirs by the Advanced Spaceborne Thermal Emission Reflection Radiometer (ASTER) and the MODerate-resolution Imaging Spectroradiometer (MODIS), LST products from the two sensors were examined for a portion of suburb area in Beijing, China. We selected Soil-Adjusted Vegetation Index (SAVI), Normalized Multi-band Drought Index (NMDI), Normalized Difference Built-up Index (NDBI) and Normalized Difference Water Index (NDWI) as representative remote sensing indices for four land cover types (vegetation, bare soil, impervious and water area), respectively. By using support vector machines, the overall classification accuracy of the four land cover types with inputs of the four remote sensing indices, extracted from ASTER visible near infrared (VNIR) bands and shortwave infrared (SWIR) bands, reached 97.66%, and Kappa coefficient was 0.9632. In order to lower the subpixel temperature estimation error caused by re-sampling of remote sensing data, a disaggregation method for subpixel temperature using the remote sensing endmember index based technique (DisEMI) was established in this study. Firstly, the area ratios and statistical information of endmember remote sensing indices were calculated from ASTER VNIR/SWIR data at 990 m and 90 m resolutions, respectively. Secondly, the relationship between the 990 m resolution MODIS LST and the corresponding input parameters (area ratios and endmember indices at the 990 m resolution) was trained by a genetic algorithm and self-organizing feature map artificial neural network (GA-SOFM-ANN). Finally, the trained models were employed to estimate the 90 m resolution subpixel temperature with inputs of area ratios and endmember indices at the 90 m resolution. ASTER LST product was used for verifying the estimated subpixel temperature, and the verified results indicate that the estimated temperature distribution was basically consistent with that of ASTER LST product. A better agreement was found between temperatures derived by our proposed method (DisEMI) and the ASTER 90 m data (R2 = 0.709 and RMSE = 2.702 K).

Highlights

► Subpixel land surface temperature was estimated from solar-reflective and thermal infrared remote sensing data. ► A novel endmember index based techniques was employed. ► The optimized neural networks improved the estimation accuracy. ► A case examination on ASTER and MODIS products over a heterogeneous area were verified. ► The proposed approach could be used to obtain finer resolution temperature.

Introduction

When referring to land observation from space, the land surface temperature (LST) is a physical parameter very important to a wide variety of scientific studies (Sobrino et al., 1991, Wan and Li, 1997, Qin et al., 2001). Satellite observations of LST have been widely applied in environmental monitoring, such as detecting conditions prone occurring wildfire (e.g., Pu et al., 2007), assessing ecosystem health and drought severity (e.g., Kustas & Anderson, 2009), monitoring volcanic eruptive activity (e.g., Lombardo & Buongiorno, 2006), and exploring urban heat island effects (e.g., Weng et al., 2004, Stathopoulou and Cartalisa, 2009). A common use of thermal infrared (TIR) data is to derive surface energy budgets (Kustas et al., 2003, Liang, 2004) from high-resolution thermal data, providing assessments of evapotranspiration (ET) down to scales of individual agricultural fields (Loheide & Gorelick, 2005). This type of thermal information is necessary for a reasonable allocation of water resources in northwest China as well as in other arid and semiarid regions around the world. However, due to a relatively lower thermal radiation that is emitted by land surfaces, most satellite sensors are not capable of providing as much finer-scale information in thermal bands as in visible and short infrared ones. Therefore, these satellite-based thermal observations generally reflect a tradeoff between temporal and spatial resolutions such that the thermal systems have either high-spatial/low-temporal resolutions (e.g., Advanced Spaceborne Thermal Emission and Reflection Radiometer-ASTER) or low-spatial/high-temporal resolutions (e.g., Terra/Aqua -Moderate Resolution Imaging Spectrometer-MODIS). Moreover, due to the heterogeneous nature of land surfaces, mixed pixels of multiple anisothermal objects exist in TIR data at a relatively coarse spatial resolution. As a result, LST products retrieved from satellite TIR data are often composed of a mixture of different temperature components (Lakshmi and Zehrfuhs, 2002, Liu et al., 2006). Therefore, there is a need for enhancing the spatial resolution of the current LST products. In particular, as coarse spatial resolution thermal sensors, such as MODIS, provide LST products with very high temporal frequencies, improvements of the spatial resolution of such frequently used LST products will significantly extend their potential uses in many applications that require LST products with a higher spatial resolution.

In recent years, many studies have focused on the estimation of subpixel LST and developed a number of corresponding approaches. Because spatial heterogeneity is closely related to scale, its effects on the satellite-derived LST are usually addressed by a scaling approach. All scaling approaches can be categorized as statistical regression methods (e.g., Sandholt et al., 2002, Kustas et al., 2003, Agam et al., 2007, Yang et al., 2010), modulation methods (e.g., Liu and Moore, 1998, Nichol, 2009, Stathopoulou and Cartalisa, 2009) and hybrid methods (e.g., Pu et al., 2006, Liu and Pu, 2008), at either thermal digital number (DN) level, or thermal radiance level or LST level. It should be pointed out that most of such studies with the scaling approaches used simulated thermal data (i.e., from fine resolution data simulated to coarse resolution data then downscaled to fine LST data).

For statistical regressive methods, the relationship between a vegetation index/variable (e.g., the Normalized Difference Vegetation Index, NDVI; vegetation fraction,fv) and radiometric surface temperature can be addressed by a polynomial or power function. Due to the fact that vegetation index (VI) is often available at a finer pixel resolution than that of LST, there is a potential to make use of the VI-LST (satellite-derived radiometric surface temperature) relation to derive LST at the VI pixel resolution. However, based on review of existing literature, to our knowledge, many statistical regressive methods did not yield ideal results in areas where vegetation cover is mixed with other land cover types, especially for areas covered with mixture of bare soil, water and impervious components (Yang et al., 2010). In fact, there are many bare lands, lakes or rivers, and villages or towns in small scale, scattering in the agriculture fields, which is the typical feature of land cover in the north of China (Zhou et al., 2010). In order to obtain high-temporal resolution temperature and ET information of individual agricultural fields to support corps drought monitoring, the downscaling method of land surface temperature (LST) should account for the major factors of different land cover types that impact the thermal properties of pixels with mixed land cover types. Therefore, to estimate subpixel temperature from coarse resolution TIR data in a heterogeneous area, other effective satellite-based indices should be considered. A common characteristic of the statistical methods is based on such an assumption that a unique VI-LST relationship does exist within a scene at multiple spatial resolutions, largely related to fractional vegetation cover. However, for a heterogeneous area, the extracted regression relationship of finer scale cannot be directly applied in the mode of coarse spatial resolution, because there are a great amount of mixed pixels existing in the coarse resolution image. To address this issue, the endmember index (EMI) based technique was employed in this study, which will be specifically discussed in the following sections.

Modulation methods concentrate more on how to distribute the thermal radiance of a thermal pixel block and to balance it among those visible pixels in such a big block; many scaling factors, such as effective emissivity, detailed LST distribution in the same season derived from other sensors owning thermal bands with higher spatial resolutions or combining emissivity and LST altogether (Nichol, 2009), can be selected as the distribution factors. Hybrid algorithms combine statistical regression and modulation methods, while straightforwardly connecting subpixel information with thermal block pixel radiance on the basis of the scale invariance of radiance. Specially, thermal radiance is linearly correlated with effective emissivity and varieties of land cover patterns that are weighted by their fractions (Liu & Pu, 2008). However, as aforementioned, the modulation methods and hybrid algorithms are generally used with the isothermal assumption under coarse resolution to retrieve the fraction and the emissivity of components. After that, they were applied to obtain subpixel temperature by adding a simple bias correction, which may cause some errors especially in vegetation area composed of a mixture of different temperature components. However, our proposed DisEMI technique, addressed below, can be served as a valuable approach when it is applied to mixed pixels to obtain different subpixel temperature.

Based on our knowledge, so far, we did not find studies that used real coarse TIR data to downscale to a fine resolution. It is necessary to test our method based on true thermal data rather than using simulated data for downscaling TIR data. Therefore, in this research, we propose a new method, which is using the endmember index based technique (including the area ratio of four land cover types, the satellite-based spectral indices, and the statistical information of the indices, e.g., average, standard deviation), the genetic algorithm and self-organizing feature map artificial neural network (GA-SOFM-ANN), to estimate 90 m resolution subpixel temperature (Ts90) from 990 m resolution MODIS LST product (Ts990) and 30 m resolution ASTER reflectance data (VNIR30). More specific objectives in this study include (1) presenting a new method we developed for estimating subpixel LST by using endmember index based technique; (2) testing the method in heterogeneous areas with MODIS and ASTER temperature products; and 3) discussing some potential problems associated with the new method.

Section snippets

Study area

Changping District, located in the northeast of Beijing City, China, was chosen as the study area (Fig. 1). The study area covers an area of 459 km2 with latitude of 40°05′20″N–40°13′50″N and longitude of 116°08′00″E–116°27′30″E. The mean elevation of study area is 42 m, and the mean terrain slope is 10 3. The representative land use/land cover (LULC) types in the study area typically consist of the vegetation, bare soil, impervious (including villages and towns), and water. More than 50% of the

Methodology

In the estimation of 90 m resolution sub-pixel temperature by using ASTER VNIR/SWIR 30 m data and the MODIS 990 m LST data, there exists a significant difference in spatial resolution. Traditionally, the ASTER data is first resampled to 990 m resolution. Secondly, the corresponding remote sensing indices are calculated for different types of land cover. Thirdly, neural networks are trained with inputs of the derived remote sensing indices and MODIS 990 m temperature data. Finally, the trained

Land cover classification map

In the study area, the derived remote sensing indices corresponding to the four land cover types were used to conduct an SVM classification. Fig. 4 shows the results of the SVM classification based on the four remote sensing indices (SAVI, NMDI, NDBI and NDWI) with training samples including 4031 pixels for vegetation, 2044 pixels for bare soil, 1034 pixels for impervious surface, and 637 pixels for water surface. Based on the classification results in the study area, the area proportions of

Discussion

Existing coarse-resolution pixel temperature downscaling methods are mostly designed for the same land cover type (mainly vegetation type). With those methods, since analysis and verification is conducted by simulating temperature data at different spatial resolutions by using a re-sampling and aggregated method, the estimated downscaling pixel temperature often has a relatively high accuracy. However, the land cover types in our experimental area in this study were non-uniform, consisting of

Summary and conclusions

As a key surface parameter, LST is of great concern in environmental studies. In this study, firstly, we used ASTER VNIR/SWIR and MODIS real remote sensing data to calculate the area ratios of the four land cover types and statistical information of endmember remote sensing indices at 990 m and 90 m resolutions. Next, a nonlinear model between pixel temperature, area ratios, and statistical information of endmember remote sensing indices at the 990 m resolution was established using GA-SOFM neural

Acknowledgements

This study was supported by the Natural Science Foundation of China (NSFC, 40901173, 41071228), Beijing Municipal Natural Science Foundation (4102021), China's Special Funds for Major State Basic Research Project (No. 2007CB714401) and Open funds of State Key Laboratory of Remote Sensing Science (2009KFJJ020), which was jointly sponsored by the Institute of Remote Sensing Applications of Chinese Academy of Sciences and Beijing Normal University. We are grateful to the anonymous reviewers for

References (47)

  • J.A. Sobrino et al.

    Atmospheric correction for land surface temperature using NOAA-11 AVHRR channels 4 and 5

    Remote Sensing of Environment

    (1991)
  • M. Stathopoulou et al.

    Downscaling AVHRR land surface temperatures for improved surface urban heat island intensity estimation

    Remote Sensing of Environment

    (2009)
  • Z. Wan et al.

    Validation of the land-surface temperature products retrieved from Terra Moderate Resolution Imaging Spectroradiometer data

    Remote Sensing of Environment

    (2002)
  • Q.H. Weng et al.

    Estimation of land surface temperature-vegetation abundance relationship for urban heat island studies

    Remote Sensing of Environment

    (2004)
  • R.E. Wolfe et al.

    Achieving sub-pixel geolocation accuracy in support of MODIS land science

    Remote Sensing of Environment

    (2002)
  • F. Yuan et al.

    Comparison of impervious surface area and normalized difference vegetation index as indicators of surface urban heat island effects in Landsat imagery

    Remote Sensing of Environment

    (2007)
  • H. Bagan et al.

    The self-organizing neural network classification study of ASTER data

    Advance in Earth Sciences

    (2003)
  • Y. Chemin et al.

    Real coded genetic algorithm for assimilating remotely sensed evapotransiration data using a soil–water–atmosphere plant model

    (2004)
  • H. Demuth et al.

    Neural Network Toolbox User's Guide, Version 4

    (2003)
  • ERSDAC. (2005). ASTER product guide: Release note. Earth Remote Sensing Data Analysis Center. Available at...
  • J. Gareth et al.

    Development and validation of a genetic algorithm for flexible docking

    Journal of Molecular Biology

    (1997)
  • A. Gillespie et al.

    A temperature and emissivity separation algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) images

    IEEE Transactions on Geoscience and Remote Sensing

    (1998)
  • S.J. Goetz

    Multisensor analysis of NDVI, surface temperature and biophysical variables at a mixed grassland site

    International Journal of Remote Sensing

    (1997)
  • Cited by (87)

    • A physical method for downscaling land surface temperatures using surface energy balance theory

      2023, Remote Sensing of Environment
      Citation Excerpt :

      However, because the LST is influenced by multiple factors, the relationships between VIs and LSTs display great limitations for heterogeneous underlying surfaces (Inamdar and French, 2009; Nichol, 2009; Stathopoulou and Cartalis, 2009; Zakšek and Oštir, 2012). To compensate for the deficiency of using only simple vegetation parameters in LST downscaling, spectral indices representing different types of land surfaces, such as the normalized difference water index (NDWI), the normalized difference built-up index (NDBI), the enhanced built-up and bareness index (EBBI), the bare soil index (BI), and the temperature vegetation dryness index (TVDI), have also been introduced (Yang et al., 2011; Zakšek and Oštir, 2012; Merlin et al., 2012; Bonafoni, 2016; Li et al., 2019; Agathangelidis and Cartalis, 2019; Liu et al., 2020). In addition, to modulate the land surface energy distribution, topographic variables, namely, digital elevation models (DEMs) and slope angle, surface emissivity and broadband albedo, were also suggested to be good scaling factors (Duan and Li, 2016; Li et al., 2019; Agathangelidis and Cartalis, 2019; Inamdar and French, 2009; Nichol, 2009; Stathopoulou and Cartalis, 2009; Zakšek and Oštir, 2012; Merlin et al., 2010, 2012; Dominguez et al., 2011).

    View all citing articles on Scopus
    View full text