Elsevier

Remote Sensing of Environment

Volume 124, September 2012, Pages 108-121
Remote Sensing of Environment

Estimating air surface temperature in Portugal using MODIS LST data

https://doi.org/10.1016/j.rse.2012.04.024Get rights and content

Abstract

Air surface temperature (Tair) is an important parameter for a wide range of applications such as vector-borne disease bionomics, hydrology and climate change studies. Air temperature data is usually obtained from measurements made in meteorological stations, providing only limited information about spatial patterns over wide areas. The use of remote sensing data can help overcome this problem, particularly in areas with low station density, having the potential to improve the estimation of Tair at both regional and global scales. Some studies have tried to derive maximum (Tmax), minimum (Tmin) and average air temperature (Tavg) using different methods, with variable estimation accuracy; errors generally fall in the 2–3 °C range while the level of precision generally considered as accurate is 1–2 °C. The main objective of this study was to accurately estimate Tmax, Tmin and Tavg for a 10 year period based on remote sensing—Land Surface Temperature (LST) data obtained from MODIS—and auxiliary data using a statistical approach. An optimization procedure with a mixed bootstrap and jackknife resampling was employed. The statistical models estimated Tavg with a MEF (Model Efficiency Index) of 0.941 and a RMSE of 1.33 °C. Regarding Tmax and Tmin, the best MEF achieved was 0.919 and 0.871, respectively, with a 1.83 and 1.74 °C RMSE. The developed datasets provided weekly 1 km estimations and accurately described both the intra and inter annual temporal and spatial patterns of Tair. Potential sources of uncertainty and error were also analyzed and identified. The most promising developments were proposed with the aim of developing accurate Tair estimations at a larger scale in the future.

Highlights

► A 1 km–8 day air T2m dataset was developed using MODIS LST and auxiliary data. ► Statistical optimization method with combined jackknife and bootstrap resampling. ► Average T2m was estimated with a global r2 = 0.94 and RMSE = 1.33 °C. ► Maximum and minimum T2m were estimated with a 1.83 and 1.74 °C RMSE, respectively. ► Main potential future developments for wider scale application were identified.

Introduction

Air temperature, typically measured at the shelter height 2 m above ground (hereafter Tair), is a key variable in a wide range of environmental applications including vector-borne disease bionomics (e.g. Kuhn et al., 2002), terrestrial hydrology (e.g. Chow et al., 1988), biosphere processes (e.g. Prince & Goward, 1995) and climate change (e.g. IPCC, 2007).

Depending on the scale, the spatiotemporal Tair patterns can be highly variable and complex due to the heterogeneity of the environmental factors that control the energy balance of the land–atmosphere system. Solar radiation is the main energy influx and the total amount reaching the Earth's surface is determined by factors, such as, (i) latitude, which determines the relative position of the sun influencing day length, thus, the distribution of total incoming solar radiation throughout the year; (ii) cloud cover and (iii) particulate matter in the atmosphere (e.g. Jacobson, 2000). The energy of the earth–atmosphere system is balanced by the absorption of incoming solar radiation, in the shortwave part of the light spectrum, and the emission of infrared long‐wave radiation and the sensible and latent heat loss fluxes (Jin and Dickinson, 2010, Prihodko and Goward, 1997). These processes impose the surface heating and cooling process which are the main modulators of the Tair daily cycle (Ahrens, 2003).

Meteorological measurements provide accurate temporally discrete Tair information but have limited ability to describe its spatial heterogeneity over large areas of the Earth. Tair measurements are frequently interpolated with significant errors associated and often lead to unrepresentative spatial patterns (Willmott & Robeson, 1995). Furthermore, the complexity associated with the correct estimation of Tair patterns increases with increased temporal resolution and when temperature extremes are the target objectives (Geiger, 1965, Vogt et al., 1997). Regardless of the method, interpolation accuracy is highly dependent on station network density and the scale of spatial and temporal variability of the parameter (Vancutsem et al., 2010, Vogt et al., 1997). Interpolation errors generally range from 1 to 3 K depending on the spatial and temporal scale, the temperature parameter and the techniques employed (Anderson, 2002, Mostovoy et al., 2006, Vogt et al., 1997).

The use of remote sensing data can greatly improve the estimation of Tair spatiotemporal patterns thus improving the knowledge of both climate and terrestrial biological processes at regional and global spatial scales (Prihodko & Goward, 1997). In regions of the world where the density of meteorological stations is sparse and the data is unavailable or expensive, remote sensing can be an important and valuable source of information (Czajkowski et al., 2000). At a longer temporal scale, it can help to reduce the necessity of site level measurements.

Several authors have proposed methods to estimate Tair using remote sensing data. The signal acquired by satellite sensors is composed by energy emitted from the Earth's surface, energy absorbed by the atmosphere, primarily by water vapor, or re-emitted from the atmosphere (Cresswell et al., 1999, Czajkowski et al., 2000). At the surface level, this energy is a mixture of temperature from vegetation, sunlit and shadowed soil, and artificial objects, among others. Land Surface Temperature (LST) has been frequently used. The quality of LST retrievals is influenced by sensor characteristics, atmospheric conditions, variations in spectral emissivity, surface type heterogeneity, soil moisture, visualization geometry, and assumptions related to the split‐window method (Czajkowski et al., 2000, Vogt et al., 1997).

Surface temperatures are governed by land–atmosphere interactions and the energy fluxes between both. The surface energy balance is governed by down and upward radiation fluxes, and, latent and sensible heat loss fluxes. The downward flux is determined by the factors mentioned previously and by the surface's albedo which determines the fraction of total radiation reflected and absorbed by the surface. The upward flux follows the Stefan–Boltzmann expression and is determined by the surface's temperature and emissivity, thus, LST is an indicative variable of the net surface energy balance driven by long‐wave radiation surface emission. The latent and sensible heat fluxes are strongly influenced by surface temperature and the apportionment of energy between them is governed by moisture content, surface type, wind velocity and emissivity (Jacobson, 2000, Monteith, 1981). It is commonly assumed that the variability in the land surface response forces the atmosphere greatly, influencing the Tair variability due to changes in LST, particularly in clear-sky days. In such conditions it is expectable that accurate mapping of Tair distribution is possible (Vogt et al., 1997).

Although LST and Tair are strongly correlated, both have different physical meanings, magnitudes, measurement techniques, response to atmospheric conditions and diurnal phase (Jin & Dickinson, 2010). The lapse rate between LST and Tair is controlled by a complex surface energy balance dependent on information not always provided by satellites (Prince et al., 1998, Vancutsem et al., 2010). The temperature lapse rate can vary greatly in a diurnal cycle, and patterns are influenced seasonally through variations in the day and night lengths (Zaksek & Schroedter-Homscheidt, 2009). During daytime, surface temperature is generally higher than air temperature, and at nighttime, the opposite occurs, leading to over and underestimation of air temperature, respectively (Cresswell et al., 1999). During daytime, a large fraction of energy is due to re-emitted long-wave radiation driven by solar insolation. During nighttime, Tair estimation is considered simpler due to the lack of the solar radiation effect on the thermal infrared signal (Vancutsem et al., 2010). Many other factors interplay and have impact on the LST–Tair lapse rate. Huband and Monteith, 1986a, Huband and Monteith, 1986b showed the importance of soil emissivity, moisture content and wind velocity for a good relation between surface temperature and the latent and sensible heat fluxes. The same authors stated that the variation of Tair is also controlled by advection, adiabatic processes, turbulence, and thermodynamic phase transformation. Jin and Dickinson (2010) showed the importance of cloud cover, water vapor content and vegetation on the land–atmosphere system. Additionally, elevation, topography and surface roughness are also important factors (e.g. Huband and Monteith, 1986a, Mildrexler et al., 2011).

Vegetation plays an important role on the LST–Tair dynamics due to exchanges of water and energy between land and atmosphere, partitioning the incoming energy in latent and sensible heat (Goward et al., 2002, Mildrexler et al., 2011; Stisen et al., 2007). Vegetation exchanges absorbed solar radiation through evapotranspiration, leading to transpirational cooling and larger latent heat fluxes. Vegetation also influences the energy balance through their low albedo and large surface roughness which promotes efficient sensible heat dissipation (Lee et al., 2011, Mildrexler et al., 2011). Additionally, the contributions of soil evaporation and transpiration to surface latent heat flux are related to changes in the leaf area index (LAI) (Nemani et al., 1996).

Zaksek and Schroedter-Homscheidt (2009) reviewed the types of methods commonly used to estimate Tair based on LST, dividing them in three distinct groups:

  • 1)

    Statistical approaches based on regression techniques which can be simple if only based on LST and Tair (e.g. Mostovoy et al., 2006, Vogt et al., 1997) or advanced, when more than one independent variable is used (Cresswell et al., 1999, Jang et al., 2004). Statistical methods generally perform well, within the spatial and time frame they were derived, but have limited generalization and require large amounts of data to train the algorithms (Stisen et al., 2007);

  • 2)

    The temperature–vegetation index (TVX) is based on the assumption that for an infinitely thick canopy, the top-of-canopy temperature is the same as within the canopy (Czajkowski et al., 2000, Prihodko and Goward, 1997) and uses the Normalized Difference Vegetation Index (NDVI) as a key input variable. However, the assumption of linear and negative slope between LST and NDVI is not always applicable and is influenced by seasonality, ecosystem type and soil moisture variability (Sandholt et al., 2002, Vancutsem et al., 2010);

  • 3)

    Energy-balance approaches are physically based. The sum of incoming net radiation and anthropogenic heat fluxes is considered equal to the sum of the surface's sensible and latent heat fluxes (Meteotest, 2010, Sun et al., 2005). The major drawback of these methods is that they require large amounts of information often not provided by remote sensing (Mostovoy et al., 2006, Prince et al., 1998).

Most of the previous studies have focused on estimating daily or instantaneous Tair. The TVX method has been widely used for Tair estimation. Czajkowski et al. (2000) estimated Tavg for a weekly period with associated RMSE between 1.72 and 3.48 °C and r2 = 0.64. Stisen et al. (2007) and Prihodko and Goward (1997) estimated Tair with RMSE higher than 2.5 °C and r2 between 0.64 and 0.86. Cresswell et al. (1999) used a statistical method to derive instantaneous Tair with an associated RMSE below 3 °C for more than 70% of sampled data. Zaksek and Schroedter-Homscheidt (2009) used a more sophisticated method based on the energy balance to estimate instantaneous Tair with an RMSE of 2 °C. Vancutsem et al. (2010) used 1 km MODIS data to estimate weekly Tmin and Tmax. They reported correlations between LST and Tmin ranging from 0.01 to 0.96 for several stations and Tmax was estimated with an r2 = 0.92 and RMSE = 1.83 °C. In sum, despite of the methods used, previous studies reported errors of about 2–3 °C for a variety of target variables and both spatial and temporal resolutions (Zaksek & Schroedter-Homscheidt, 2009). The level of precision generally accepted as ‘accurate’ for remote sensing based Tair estimations is between 1 and 2 °C (Vazquez et al., 1997).

The remote sensing based Tair estimations are usually based on models that are designed to infer from sparse data, measured in meteorological stations, the Tair distribution over larger areas. The measured data set is a subset of the overall distribution of Tair thus it contains partial information about the ‘reality’. It is important to potentiate the information contained in model predictions and reduce the biased view of ‘reality’ created by using a specific training data set (Rogers, 2006). Statistical resampling methods have been used to create multiple approximate distributions based on the training data set, allow attaining robust and representative distributions of model accuracy, and derive uncertainty estimations along with confidence intervals for model predictions. Moreover, they permit to obtain distributions for all model parameters, maximizing the confidence of future extrapolations while minimizing the influence of smaller sub sets (e.g. outliers) in the model calibration.

The main objective of the study presented in this paper was to estimate average, minimum and maximum Tair (hereafter Tavg, Tmin and Tmax respectively) based on remote sensing and auxiliary data using an advanced statistical approach. Detailed analysis was focused on Tavg. Typically, previous studies which aimed at estimating Tair using remote sensing data have limited temporal and spatial coverage ranging from a small number of days to a few years and generally applied to small areas and/or covered by a small number of meteorological stations (e.g. Cresswell et al., 1999, Stisen et al., 2007, Vogt et al., 1997). We estimated Tair for the whole Portuguese mainland territory for a complete decade using 106 stations. Moreover, the majority of studies estimate Tair at short temporal scales, ranging from hourly to daily resolutions (e.g. Mostovoy et al., 2006, Stisen et al., 2007). In this study we estimate weekly Tair and aim at describing accurately its intra and inter annual patterns, which is the relevant scale for a wide scope of environment related applications. Furthermore, to the authors’ knowledge, despite the numerous advantages of coupling resampling methods to modeling exercises its application to Tair estimation based on remote sensing data has seldom been made. In the present work, resampling methods were implemented to maximize the information of the sparse station level Tair data and model predictions, in order to increase the accuracy and confidence of estimations. The errors associated with Tair estimation based on remote sensing data are often large and strongly limit its applicability (e.g. Czajkowski et al., 2000, Vazquez et al., 1997, Vogt et al., 1997). One of the objectives of this work is to provide Tair estimations with an accuracy which will potentiate the future applications, in agreement with the 1–2 °C interval suggested by Vazquez et al. (1997). The main factors and processes influencing the Tavg estimation errors are also explored. Finally, this work highlights its limitations and possible future developments.

Section snippets

Meteorological station data and study region

The Iberian Climate Atlas (AEMET & IM, 2011) shows that annual average temperature in continental Portugal varies between 10 and 20 °C, with strong spatial gradients: lower temperatures are associated with regions over northern latitudes, closer to the coast, and with higher altitudes. The seasonal variation is between cold to mild winters (minimum monthly Tavg between 0 and 12 °C) and mild to warm summers (maximum monthly Tavg15 and 27.5 °C), with a spatial distribution similar to that of annual

Overall performance

Due to the large amount of data (N = 34117) and the resampling methods used, the calibration and validation accuracy were identical (not shown). Several models were implemented but for simplicity only a subset of the models used for Tair estimation are shown in Table 1 along with their respective validation performance statistics. The MEF equaled the r2 statistic for all models listed in Table 1.

A simple linear equation using only daytime LST explained 83.3% of the variability, while using only

Air temperature estimation

The results of this study can be considered promising, given the simplicity of the statistical models employed, the robustness of the jackknife and bootstrap techniques and the high accuracy achieved. To support these findings, for Tavg, 85% of the stations had an r2 higher than 0.90 and RMSE lower than 1.5 °C. The LST Night and LST Day were the most relevant predictors explaining, respectively, between 64–87% and 85–92% of the variability of the three Tair variables (Table 1). This result shows

Conclusions

This work showed that weekly average Tmax, Tmin and Tavg can be accurately estimated using remote sensing techniques. The simple statistical models employed estimated Tavg with a MEF of 0.941 and a RMSE of 1.33 °C. Regarding the Tmax and Tmin, the best MEF achieved were 0.919 and 0.871, respectively, with a 1.83 and 1.74 °C RMSE. The developed datasets accurately described both the intra-inter annual and spatial patterns of Tair. The model had a general tendency to have lower performance in

Acknowledgments

This work was funded by the Portuguese Foundation for Science and Technology through the research project MALVEO (contract no. PTDC/CLI/67910/2006) and the post-doctoral fellowship attributed to J.P. Nunes (ref. SFRH/BPD/39721/2007). We also acknowledge the measured air temperature data providers, namely, SNIRH, Univ. Aveiro (esp. Dr. Jacob Keizer), Univ. Évora and the WMO. We acknowledge the MODIS mission scientists and associated NASA personnel for the production of the data used in this

References (69)

  • C. Vancutsem et al.

    Evaluation of MODIS land surface temperature data to estimate air temperature in different ecosystems over Africa

    Remote Sensing of Environment

    (2010)
  • Z. Wan

    New refinements and validation of the MODIS land-surface temperature/emissivity products

    Remote Sensing of Environment

    (2008)
  • Z.M. Wan et al.

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

    Remote Sensing of Environment

    (2002)
  • K. Zaksek et al.

    Parameterization of air temperature in high temporal and spatial resolution from a combination of the SEVIRI and MODIS instruments

    ISPRS Journal of Photogrammetry and Remote Sensing

    (2009)
  • S. Ackerman et al.

    Discriminating clear sky from clouds with MODIS

    Journal Geophysical Research

    (1998)
  • AEMET et al.

    Iberian climate atlas. Air temperature and precipitation (1971–2000)

    AEMET: Agencía Estatal de Meteorología and IM: Instituto de Metereologia

    (2011)
  • C.D. Ahrens

    An introduction to weather, climate, and the environment

  • S. Anderson

    An evaluation of spatial interpolation methods on air temperature in Phoenix, Arizona State

    (2002)
  • D.G. Baker

    Effect of observation time on mean temperature estimation

    Journal of Applied Meteorology

    (1975)
  • M. Bossard et al.

    CORINE land cover technical guide—Addendum 2000

    Technical Report No. 40

    (2000)
  • C. Cerdeira et al.

    Validação das cartografias globais de ocupação do solo, GLC2000 e MOD12Q1, para Portugal Continental

    Proceedings of ESIG 2006 - IX Congress of Geographic Information Users

    (2006)
  • V.T. Chow et al.

    Applied hydrology

    (1988)
  • T.F. Coleman et al.

    On the convergence of reflective newton methods for large-scale nonlinear minimization subject to bounds

    Mathematical Programming

    (1994)
  • M.P. Cresswell et al.

    Estimating surface air temperatures, from Meteosat land surface temperatures, using an empirical solar zenith angle model

    International Journal of Remote Sensing

    (1999)
  • K.P. Czajkowski et al.

    Thermal remote sensing of near surface environmental variables: Application over the Oklahoma Mesonet

    The Professional Geographer

    (2000)
  • A.C. Davison et al.

    Bootstrap methods and their application

  • B. Efron

    Bootstrap methods—another look at the Jackknife

    The Annals of Statistics

    (1979)
  • R. Geiger

    Climate near the ground

    (1965)
  • N.D.S. Huband et al.

    Radiative surface temperature and energy balance of a wheat canopy I: Comparison of radiative and aerodynamic canopy temperature

    Boundary-Layer Meteorology.

    (1986)
  • N.D.S. Huband et al.

    Radiative surface temperature and energy balance of a wheat canopy II: Estimating fluxes of sensible and latent heat

    Boundary-Layer Meteorology

    (1986)
  • IPCC

    Climate Change 2007: The Physical Science Basis. Contribution of Working Group I to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change

  • M. Jacobson

    Fundamentals of atmospheric modeling

    (2000)
  • J.D. Jang et al.

    Neural network estimation of air temperatures from AVHRR data

    International Journal of Remote Sensing

    (2004)
  • A. Jarvis et al.

    Hole-filled seamless SRTM data V4

    (2008)
  • Cited by (367)

    View all citing articles on Scopus
    View full text