Open Access
Add to BibSonomy
Abstract
Land surface temperature (LST) is a key parameter in the interaction of the land-atmosphere system. Nevertheless, on the regional scale, the actual weather is cloudy for half a year in most regions. Therefore, receiving all-weather LST from thermal-infrared remote sensing is necessary and urgent. In this paper, an approach with multi-temporal and spatial neighboring-pixels in combination with diurnal solar radiation and surface temperature evolution is proposed to estimate daytime all-weather LST using FY-2D data. Evaluation of the accuracy of the algorithm is performed against the simulated data and the in situ measurements. The root mean square error (RMSE) between the actual and estimated LSTs under cloud-free conditions is approximately 1.84 K for the simulated data, while the RMSE of LST under cloud-free conditions varies from 3.42 to 5.1 K for the in situ measurement, and RMSE of LST under cloudy sky is approximately 7 K. The results indicate that the new algorithm is practical for retrieving the daytime all-weather LST at high-temporal resolution without any auxiliary field measurement, although some uncertainties exist.
© 2017 Optical Society of America
1. Introduction
Land surface temperature (LST) is one of the most important variables in monitoring surface energy and global climate change, and it controls many biological and physical processes between atmosphere and land [1, 2].
Due to the characteristic of the remote sensing, such as simultaneous observation, speediness and comprehensiveness, the technology is the only means to observe LST over the entire globe with some temporal resolution and spatial coverage. Because of the simplicity and practicality of the general split window method (GSW), currently, LST is obtained from thermal infrared remote sensing data. But the method is only used to estimate LST under the cloud-free condition (Tclear) and many products only supply the LST for clear pixels while invalid value was filled for cloudy pixels. However, on the regional scale, the actual weather is cloudy for half a year in most regions [3]. Therefore, how to get all-weather LST is necessary and urgent.
At present, some methods have been developed to estimate all-weather daytime LST. Due to the penetrability of the microwave, LST under the cloudy-sky condition (Tcloud) is retrieved using the data. Many researchers proposed some methods to estimate LST, especially under the cloudy-sky condition [4], but the accuracy is poor. To be specific, most results show that the root mean square errors (RMSE) ranged from approximately 2.9 K to 8.5 K during daytime and from approximately 2.1 K to 3.3 K during nighttime. Meanwhile, the spatial resolution of subsurface temperature from microwave data is low (e.g., 25 km for AMSR-E). Therefore, an original method of retrieval of all-weather LST combing the polar-orbiting thermal infrared and passive microwave data is proposed [3]. The method provided all-weather LST at high-spatial resolution rather than at high-temporal resolution.
Up to now, all-weather LST is often retrieved using thermal infrared data by the reconstruction technology, which is effectively recovering the missing information, and the reconstruction is realized using spatial or temporal information. Traditional methods using spatial information include inverse distance weighting, spline function and geo-statistical interpolation methods, like co-kriging [5,6]. Because of the limited spatial information, some researchers have attempted to introduce more information in reconstructing LST such as elevation, temporal information [7]. Zeng et al. [8] filled invalid LST values using a multi-temporal classification and a robust temporal regression, and the method can fill invalid LST values accurately with much less reference data. However, the aforementioned methods did not consider the character of cloud and only provide hypothetic clear-sky LST.
In order to obtain the real LST under the cloudy-sky condition, Jin et al. [9] proposed a ‘spatial neighboring-pixel’ approach to estimate the Tcloud from polar-orbiting satellite data, in which Tcloud is interpolated from LST observations of surrounding Tclear pixels within 100 to 300 km or within two days based on the surface energy balance. However, the method is limited if the clear and cloudy pixels are not homogeneous or the atmospheric conditions are non-uniform. To overcome this deficiency, a method is proposed using temporal-based Tclear to estimate Tcloud [10]. At the same time, the result is also compared with the LST estimated using the spatial-based neighboring-pixel method. The result shows that the temporal ‘neighboring-pixel’ method is better than a spatial approach, and the absolute error is within 1.5 K. However, some disadvantages of this approach are inevitable. First, the method is proposed based on the same or similar Tclear of temporal neighboring-pixel. In fact, because the difference of Tclear at two times is obvious especially for the longer time interval, large estimation errors of Tcloud will be produced if the time interval is longer between the time that cloud appears and that the sky is clear. Second, some errors can be caused due to the inconsistence that Tcloud is interpolated from Tclear of temporally neighboring pixels, whereas the difference in net solar shortwave radiation (NSSR) is obtained from spatially neighboring pixels. Third, auxiliary field measurements like solar radiation need to be input in the method. To avoid the weakness, Yu et al.[11] proposed a spatially and temporally neighboring-pixel method to reconstruct cloud-contaminated pixels using daily MODIS LST products based on the consideration of the surface energy balance. Nevertheless, ground-based measurements are often needed to calculate Tcloud, which makes it difficult to implement for the fewer observation regions. In order to reduce the input of the in situ measurement data, Zhang et al. [12] proposed a method to estimate Tcloud based on a one-dimensional heat transfer equation and the evolution of daily temperatures and NSSR. The method only uses multi-temporal satellite data to obtain Tcloud, but it is invalid when Tclear on one day is less than six observations.
To reduce the dependence on auxiliary field measurements, this study aims to develop an improved flexible and effective method to retrieve all-weather daytime LST. The advantage of the improved approach is that it can generate more accurate LST for all the pixels, without depending on ground-based ancillary data and the amount of Tclear on one day and the method can provide LST at high-temporal resolution. This approach could be valuable in meteorological and hydrological studies and applications. In the following sections, the methodology is firstly put forward in section 2; next, the data including simulated data, satellite data and field measurement are described in section 3; then, the method’s performance is demonstrated based on both simulated and field measurement in section 4; Finally, the conclusion is drawn in section 5.
2. Methodology
All-weather daytime LST is retrieved with two steps: one step is to estimate the Tclear using GSW method; the other step is to estimate the Tcloud using multi-temporal pixel or spatial neighboring-pixel method in combination with the diurnal solar radiation and the surface temperature evolution. Figure 1 shows the flowchart of the retrieval of all-weather LST using FY-2D data.
2.1. Estimation of LST under cloud-free conditions
Due to the simplicity and operability of GSW method, LST is retrieved using the GSW method for most satellite data including polar-orbiting and geostationary satellite [13]. Considering that onboard S-VISSR sensor of FY-2D has two adjacent thermal infra-red channels, LST under the cloud-free condition is estimated using GSW algorithm proposed by Sobrino et al. [14]. The specific formula can be expressed as following:
where Ts is LST, Ti and Tj are the top of the atmosphere (TOA) brightness temperatures measured in channels i (11.0µm) and j (12.0 µm), respectively, Δε = (εi-εj)/2, εi and εj are the land surface emissivities (LSEs) in channels i and j, respectively. a0-a5 are unknown coefficients which will be obtained through statistical regression method from simulated data mentioned in section 3.1.2.2. Estimation of LST under cloudy-sky conditions
If the pixel is overcast by cloud, the LST cannot be obtained through GSW method. LST under the cloudy-sky condition can be derived using multi-temporal pixel method in combination with diurnal solar radiation and surface temperature evolution if the clear-sky LSTs observations on one day are more than 6 and the clear-sky LSTs observations in the morning are more than 2 [12]. However, if the conditions cannot be met, LST under cloudy-sky condition should be estimated using spatial neighboring-pixel method in combination with diurnal solar radiation and surface temperature evolution.
2.2.1. Estimation of cloudy LST using the multi-temporal pixel method
Assuming the variation in the LST is caused by variations in insolation (ΔS), Zhang et al. [12] proposed a method to calculate the LST under cloudy-sky condition.
where Tcloud is the LST under cloudy–sky condition, Tclear is the hypothetic LST under cloud-free condition at the same time, and ΔS is the difference between a hypothetic clear-sky NSSR and real NSSR under cloudy–sky condition at the same time, and P represents the resistance to a temperature change in the upper few centimeters of the surface throughout the day. Considering the lag phenomenon of variation in LST to variation in insolation, ΔS cannot reach the maximum once cloud appears, but at some time after cloud appears (td-ts), and the degree of the influence increases incrementally (from 0 to 1). So, ΔS is calculated as follows:
where Sactual is the actual NSSR at a moment, Sfit is a hypothetic clear-sky NSSR, t is the time, td is the time at which the temperature is maximized, ts is the time at which the NSSR is maximized, tnow is the calculated time.
According to Eqs. (2) and (3), Tcloud is obtained on the condition that six parameters (a hypothetic Tclear, a hypothetic clear-sky NSSR (Sfit), actual NSSR at a moment (Sactual), ts, td and P) are known. In the following section will represent the retrieval of the six variables.
NSSR under all kinds of conditions can be estimated using the method by proposed Zhang et al. [15] is shown in Eqs. (4) and (5).
where as represents the ratio of the net surface absorb radiation to the outgoing flux at the TOA; E0 is the TOA solar irradiance at one astronomical unit (W/m2), θs is SZA, d is Earth– Sun distance in astronomical units, andwithwhere r is broadband albedo at TOA. ρ1 is the reflectance of VIS band at TOA which can be obtained from FY-2D VIS channel; c0 and c1 are constants which will be obtained through statistical regression method from simulated data mentioned in section 3.1 and they are the function of VZA. μ is the cosine of VZA, wvc is total atmospheric water vertical content, b1-b7 and x, y, z are constants.Net short solar radiation (NSSR) diurnal evolution (DSC) for the clear day can be expressed as Eq. (6) [12]
where Smin approximately represents the minimum NSSR in the daytime, Smax is the amplitude of NSSR in the daytime (generally, it appears at noon), ws is the angular frequency that is nearly π/DD (DD is the duration of daytime), and ts is the time at which the NSSR is maximized (NSSRmax).NSSR at any time in one day can be obtained regarding as a hypothetic clear-sky NSSR if four parameters (Smax, Smin, ws and ts) can be estimated using four clear-sky NSSRs during the daytime with the Eq. (5).
Similarly, as is known to all, LST diurnal variation can be expressed as Eq. (7) (diurnal temperature cycle (DTC) model) [16,17,18].
where Tmin approximately represents the minimum temperature, T0 represents the amplitude (approximately Tmax-Tmin, where Tmax is the maximum LST in the daytime) and wd is the angular frequency, which is nearly π/DD (DD is the duration of daytime), td is the time at which the temperature is maximized (Tmax), trs is the starting time of the attenuation (near sunset), and β is the decay coefficient during the nighttime.
As mentioned in section 2.1, Tclear can be estimated using GSW method, and LST at any time in one day can be obtained regarding as a hypothetic Tclear if six parameters (Tmin, T0, wd, td, β and trs) can be estimated using six clear-sky LST observations during the day with the Eq. (7).
Based on the energy-balance equation and the express of ground heat transport, adopting DSC and DTC model (Eqs. (6) and (7)), at the same time, assuming (1) Latm↓-H-LE = aT + b (H represents sensible heat flux and LE represents latent heat flux); (2) the surface long-wave radiation function is linearized in the vicinity (Ti); (3) the frequency of the solar radiation and the surface temperature diurnal change are the same, Zhang et al. proposed a method to calculate thermal inertia just listed in Eq. (8) [12].
Currently, the six parameters (a hypothetic Tclear, a hypothetic clear-sky NSSR, actual NSSR at the same moment, ts, td and P) can be obtained using Eqs. (4)-(8), and Tcloud can be estimated using the method if there are more than six observations including Tclear and NSSR (at least include two observations in the morning) in one day can be obtained. If the condition cannot be met, Tcloud cannot be obtained using this method. At this time, a novel algorithm which uses spatial neighboring-pixel method to estimate Tcloud in combination with diurnal solar radiation and surface temperature evolution is proposed.
2.2.2. Estimation of cloudy LST using the spatial neighboring-pixel method
If there are less than six LST observations under a clear sky in one day or less than two clear-sky LST observations in the morning, retrievals of LST under cloudy condition need to use spatial neighboring-pixel method in combination with diurnal solar radiation and surface temperature evolution. The specific process is as follows: first, spatial neighboring-pixels which met the conditions (more than six Tclear observations in one day and more than two Tclear observations in the morning) are searched within 50*50 windows (within 250 km); then, the distance between estimated pixel and searched pixel is calculated using Eq. (9), and the nearest pixel which met the condition is selected as spatial neighboring-pixel; third, assuming that the coefficients of DTC for the estimated pixel are the same as that for the spatial neighboring-pixel, and LST at any time in one day for the estimated pixel can be obtained to be regarded as hypothetic clear-sky LST; at the same time, the coefficient in the expression of TOA solar irradiation as shown in Eq. (6), is related to the declination and longitude just as shown in Eq. (10), so, the coefficients in the DSC model fitted by the spatial neighboring-pixel clear-sky NSSR observations using Eq. (6) are adjusted as the actual coefficients of estimated pixel using the declination and longitude which is shown in Eq. (11), thereby, the hypothetic clear-sky NSSR at any time in one day for the estimated pixel can be obtained; meanwhile, assuming the P of the estimated pixel is the same as that of the spatial neighboring-pixel; last, LST can be estimated with Eq. (2).
where x and y are the spatial coordinate of a pixel, i and j are the estimated and spatial neighboring-pixel respectively.
where S is the solar constant, τ is the transmission, A is the broadband albedo, Zn is the solar zenith, λ is the latitude, δs is the solar declination, NSSR is the net shortwave solar radiation (W/m2).
where Smin_near and Smax_near represent the Smin and Smax fitted by the spatial neighboring-pixels NSSR observations under cloud-free sky, respectively, λest and δest represent the longitude and declination of estimated pixels respectively, λnear and δnear represent the longitude and declination of the spatial neighboring-pixels.
3 Data
3.1. Simulated data
For clear pixels, the popular method to estimate LST is the GSW method [13]. The coefficients of GSW should be estimated using bright temperature (BT) and corresponding LST under various kind of atmospheric and surface conditions. So far, there is no available database of field LST measurement in coincidence with FY-2D. Therefore, it is more effective to make radiative transfer simulation for wide ranges of atmospheric and surface conditions [19]. In this work, the top of the atmosphere (TOA) radiance and the atmospheric parameters (Latm↑, Latm↓, τ) ranging from 0.3 to 5 µm and 8.5 to 14.5 µm at an interval of 1 µm were simulated using atmospheric radiative transfer model (MODTRAN 5.2) under various kinds of atmospheric conditions obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis and nine types of surface cover including soil, vegetation canopy, grassland, wetland, city, desert and ocean surface, new snow and sea ice. In addition, the view zenith angle (VZA) ranging from 0° to 60° at an interval of 1°, relative azimuth angle (RAA) ranging from 0° to 120° at an interval of 60° and the solar zenith angle (SZA) ranging from 0° to 60° at an interval of 10° and atmospheric water vertical content (WVC) ranging from 0 to 6.5 g/cm2 at an interval of 1 g/cm2 are all inputted into the MODTRAN5 as input parameters. In order to make the simulation more representatives, according to the atmospheric temperature Ta in the first boundary layer, LST are varied in a wide range from Ta-5K to Ta + 15K. Moreover, according to the character of the most land covers, the averaged emissivity ranges from 0.90 to 1.0 with a step of 0.01 and the emissivity difference ranges from −0.01 to 0.01 with a step of 0.005.
Then, the TOA radiance and the atmospheric parameters (Latm↑, Latm↓, τ) corresponding to the upgraded Stretched-Visible and Infrared Spin-Scan Radiometer (S-VISSR) onboard FY-2D were obtained with the appropriate thermal infrared and VIS channel response function of the S-VISSR onboard FY-2D. Last, for given LST, in combination with the atmospheric parameters (Latm↑, Latm↓, τ) obtained from the output result of MODTRAN 5.2 and the emissivity of two thermal infrared channels, the channel brightness temperature (BT) at the TOA can be determined with the inverse of Planck’s law. So far, the database including LST and BT corresponding to different WVC, view zenith angle, solar zenith angle, emissivity, and surface condition is set up and 4755135 different situations are included in the database.
3.2. Satellite data
3.2.1. MODIS data
MODIS (Moderate Resolution Imaging Spectro-radiometer) is a key instrument carried on the Terra and Aqua satellites. Terra's orbit around the Earth is timed so that it passes from north to south across the equator in the morning, while Aqua passes south to north over the equator in the afternoon.
MODIS data products have been available and widely used in many studies and applications since 2000. In the study, the required MODIS data products include the MODIS/Aqua LST Daily L3 Global 1 km Grid product (MOD11A, Collection 6) which was used to provide LST and daily emissivity and land use product (MCD12) which is used to discriminate different land. In order to match MOD11A product and FY-2D data in a consistent system, the MODIS Re-projection Tool (MRT) is used to re-project MOD11A product and MCD12 from a sinusoidal projection to a geographical projection.
3.2.2. FY-2D data
In this study, FY-2D data downloaded from the China Meteorological website (http://satellite.cma.gov.cn/PortalSite/Data/Satellite.aspx) was used to evaluate the proposed work-frame. FY-2D, a geostationary meteorological satellite developed by Shanghai Academy of Space Flight Technology and China Academy of Space Technology was launched on 8, December 2006 and is located above the Equator at longitude 86.5° E and 35,800 km away. One of the FY-2D missions is to provide meteorological information for the Asia-Pacific region. The S-VISSR as the main sensor are loaded on the satellite, and it can acquire one full disc image covering the Earth surface from 60° N to 60° S in latitude and from 45° E to 165° E in longitude per hour and 30 min per acquisition for flood season. S-VISSR consists of 5 channels, including a VIS channel and 4 infrared channels (Table 1). Latitude, longitude, VZA, SZA, and RAA besides radiances of VIS and thermal infrared channel were also provided by the downloaded disc image file. Considering the method is different for the clear and cloud pixels, cloud products of FY-2D were also used to discriminate the different weather conditions. Meanwhile, geolocation files are used to provide the latitude and longitude for each pixel.
Table 1. Main technical index of FY2-VISSR radiometer.
3.3. Ground data
Field measurements were collected to evaluate the proposed method from Taiyuan Xiaodian meteorological station (112°33′ E, 37°47′ N) with the elevation of 780m and Changwu ecological station (107°40′E, 35°12′N) with the elevation of 1300m, which is located in Shaanxi Province and joined the Chinese Ecosystem Research Network in 1991. The two stations both belong to the same land cover type, i. e., crop land which can be found in Fig. 1 (Changwu station was dominated by the crop and Xiaodian station was dominated by the grass) and photos of two experiments station are shown in Fig. 2. In this study, LST data are detected using PTB100 and collected at 1minute intervals from March 1 to April 26, 2012. In situ LST data corresponding to the image time of FY-2D are selected for validation.
4 Results and validation
4.1 Retrieval of LST under clear conditions
As is shown in Eq. (1), LSEs are required as a model input in GSW algorithm. In the following, the method to obtain LSEs is described.
LSEs in channels IR1 and IR2 of S-VISSR can be obtained from the daily MODIS LST product (MOD11A1) at 1 km resolution. Due to the slight difference between the two split-window channels of MODIS and S-VISSR, the relationships between MODIS and S-VISSR on LSEs are proposed using the spectral databases from the Johns Hopkins University (JHU) (http://speclib.jpl.nasa.gov/), which were shown in Eqs. (12) and (13).
Figure 3 shows the LSs and linear regression results. As shown in Fig. 3, the results of the linear regression are good with the determination coefficients more than 0.998, which indicates that the LSEs in S-VISSR channels IR1 and IR2 can be directly derived from those in MODIS channels 31 and 32, respectively.Due to the difference of spatial resolution of S-VISSR and the MODIS, they should be matched accurately. First, LSEs from MOD11A1 were aggregated into WGS 84 (World Geodetic System 1984) coordinate system with the spatial resolution of 5KM using weighed averaged method. Then, LSEs were extracted using spatial nearest neighboring method according to the geolocation from FY-2D.
In order to improve the accuracy of estimation of LST, WVC is divided into some sub-ranges ranging from 0 to 6.5 g/cm2 at an interval of 1 g/cm2 while determining the a0-a5 coefficients in Eq. (1). The retrieval of WVC using three methods is compared by Zhang et al. [21] with MSG2-SEVIRI Thermal-IR data and the better result is got from split window co-variance ratio method just listed in Eq. (14) [20, 21]. So, WVC is obtained using this method in this study assuming that the ratio of two adjacent thermal infra-red channel emissitivities equals to 1.
with
where o1 and o2 can be derived as functions of VZA just shown in Eqs. (15) and (16). i and j are the split-window channels, the subscript k denotes pixel k, and the ₸ is the TOA mean (or the median) channel brightness temperatures of the N neighboring pixels considered for channels i and j, respectively.The determination of coefficients a0-a5 is performed after dividing the WVC into six sub-ranges with an overlap of 0.5 g/cm2: [0, 1.5], [1.0, 2.5], [2.0, 3.5], [3.0, 4.5], [4.0, 5.5], and [5.0, 6.5] g/cm2. Meanwhile, a0-a5 is interpolated according to θ at an interval 1°. Tclear is calculated using the GSW algorithm with the coefficients corresponding to kinds of the sub-range WVC under view zenith angle (θ equals to 0), which is shown in Fig. 4. The RMSE between the actual Tclear and estimated Tclear is from 0.74 to 1.84K with a mean error of −0.05K for different sub-ranges of WVC. The errors of LST under cloud-free sky are increasing with the increase of WVC and most scatters are concentrated on [-1, 1]. Figure 5 gives the histogram of actual and estimated LST under all kinds of conditions. The result shows that the RMSE between the actual Tclear and estimated Tclear is 1.82K with a mean error of 1.43K for all conditions. Meanwhile, the errors between the actual and estimated LST is distributed within ± 6 K. Due to errors in the parameters inputted in the algorithm cause the increasing of estimated Tclear, in the following, Tclear errors caused by the uncertainty in instrumental noises (NEΔT) and LSEs (i.e. sensitivity analysis) are performed.
According to the Eq. (1), the LST error (δLST) due to the uncertainty in (NEΔT) can be estimated by Eq. (17),
where δLST is the error of LST, δTi and δTj are the error of brightness temperature in channel i and j, respectively.
An error from −0.5 to 0.5 K at an interval of 0.1 K is respectively added to the TOA brightness temperatures Ti and Tj, the RMSE of LST under kinds of θ is obtained based on the error from Eq. (17) and results are shown in Fig. 6. The results show that 1) The RMSE of actual and estimated LST is increasing with the increase of the absolute (δTi-δTj); 2) The RMSE reached minimums when the WVC ranges from 2 to 3.5 g/cm2, whereas the RMSE reached maximums when the WVC ranges from 4 to 5.5 g/cm2, and the RMSE reached 8 K when (δTi-δTj) is ± 1 K; 3) The RMSE is all less than 4K when the WVC is less than 4.5 g/cm2; 4) The RMSE is less than 4 K when the absolute (δTi-δTj) is less 0.4 K.
Similarly, According to the Eq. (1), the LST error (δLST) due to the uncertainty in LSEs can be estimated by Eq. (18),
where δεi and δεj are the error of emissivity in channel i and j.An error from −0.01 to 0.01 K at an interval of 0.005 is respectively added to LSEs (εi and εj), the RMSE of LST is obtained based on the error from Eq. (18) and results show that RMSE of actual and estimated LST are all less than 0.5 K, and a conclusion that the sensitivity caused by the uncertainty in LSEs is unobvious can be drawn.
In addition, the in situ LST measurement at two sites (Taiyuan Xiaodian and Changwu station) is used to evaluate the accuracy of LST under the cloud-free conditions. The RMSEs of measured LST and estimated LST are 3.42 and 5.12 K with a mean error (ME) of 0.004 and 3.03 K, respectively for Taiyuan Xiaodian and Changwu station, which is shown in Figs. (7) and (8). Three reasons can explain the relatively large RMSE values associated with Tclear. One reason is due to the spatial scale inconsistency between the all-weather LST and the in situ LST measurements. The in situ LST were measured within a very small area (4m*4m), while the satellite measurement was collected in a very large region (nearly 5 km * 5 km) and reflected an integrated response over a heterogeneous area. The second reason is the uncertainty of the measurement and that of GSW algorithm caused by the errors of the input parameters. Just as shown in sensitivities analysis, the RMSE of LST reached 8K when the difference of NEΔT between IR1 and IR2 channel is ± 1 K for WVC more than 4 g/cm2. The last reason is associated with the quality of FY-2D cloud product. The misjudgment of cloud will affect the right choice of the algorithm.
Fig. 7 Comparison of LST estimated using proposed method and in situ LST from Taiyuan experiment station.
Fig. 8 Comparison of LST estimated using proposed method and in situ LST from Changwu experiment station.
4.2 Estimation of LST under cloudy conditions
Figures 7 and 8 show a comparison of the estimated daytime all-weather LST using the proposed method and the daytime all-weather LST measured at the Changwu Ecosystem experimental station and Taiyuan Xiaodian meteorological observation station from March 1 to April 26 in 2012. The scatter plot shows that the RMSE is 7.59 K and 7.87 K with a mean error of −3.84K and 3.15K under cloudy condition, respectively for Taiyuan and Changwu station. Compared to the results presented in previous studies, the error appears relative larger, such as, the errors of LST under cloudy sky obtained using spatial neighboring-pixel method proposed by Jin et al.[9] are 2K for most situations and some errors reached 8-10K. But the result is performed based on the simulated data and shows a better.
The bias of LST under the cloudy sky is mainly caused by four reasons. The first reason is the spatial scale inconsistency between the estimated LST and in situ LST measurements which the spatial resolution of 4 m for in situ LST measurement and the spatial resolution of 5 km for the estimated LST in evaluating the Tclear. The second reason is the uncertainty of some parameters (Tclear, NSSR, WVC, LSEs, θ etc.) input in the proposed algorithm. Zhang et al.[12] pointed out the 0.5 K error in Tclear leads to the RMSE of 1.44 K. As mentioned in sensitivity analysis to NEΔT in section 4.1, the RMSE of actual and estimated Tclear is increasing with the increase of the absolute NEΔT, the RMSE of 3K will be caused in estimating Tclear when the difference of NEΔT is more than 0.5K and the WVC is less than 3.5g/cm2, but the RMSE of 8K will be caused in estimating Tclear while the WVC is more than 3.5g/cm2, as a result, the errors in estimating Tcloud will be increased due to the larger errors of Tclear and 8K uncertainty of estimation of Tcloud will be yielded under some situations. The third reason is the uncertainty in the proposed algorithm, such as some assumption in the algorithm (e. g. the parameters of DTC of estimated pixel are the same as that of the spatial neighboring-pixel. In fact, due to the heterogeneity of surface and atmospheric conditions, the parameters will display difference and it will yield more uncertainty in estimating Tcloud). The last reason is associated with the quality of FY-2D cloud product. The choice of algorithm depends on the cloudy judgment and cloudy misjudgment will decrease the accuracy of Tcloud. In addition, uncertainty of field measurement also affects the result.
At the same time, all-weather LSTs ranging from 90° to 115°E in longitude and 30° to 40° N in latitude are calculated during FY-2D scanning on March 23, 2012 at UTC 1:30 and shown in Figs. 9 and 10. Figure 9 gives the LSTs under cloud-free sky and Fig. 10 displays the daytime all-weather LST. Due to the existent cloud, several areas of missing data can be found in Fig. 9 and the regions of missing data are filled in Fig. 10. A phenomenon can be found that the filled effect of the algorithm will be degraded when the area of cloud contamination is larger.
Fig. 9 Spatial distribution of the all-weather LST during FY-2D scanning on March 23, 2012 at UTC 1:30 under the cloud-free condition.
Fig. 10 Spatial distribution of the all-weather LST during FY-2D scanning on March 23, 2012 at UTC 1:30.
5 Conclusion
With increasingly more geostationary meteorological satellites in operation, it is possible and realized to obtain all-weather LST at high-temporal resolution using multi-temporal satellite data without any auxiliary field measurement. In this paper, we proposed a method to estimate daytime all-weather LST by combining diurnal solar radiation with diurnal surface temperatures under different conditions without any auxiliary field measurement. The method consists of two steps, including the estimation of LST under cloud-free sky using general split window method and the estimation of LST under cloudy sky using multi-temporal data in combination with spatial neighboring-pixel method. Data collected S-VISSR sensor aboard the FY-2D was used as an example data for input into the proposed algorithm.
Considering the retrieval of Tcloud is based on the Tclear, in this study, first, the GSW method is used to obtain Tclear, and the coefficients of the GSW for FY-2D data is calculated using simulated data obtained from atmospheric radiative transfer model (MODTRAN 5.2) under various kinds of atmospheric and surface condition. The RMSE between the actual Tclear and estimated Tclear is 1.82K with a mean error of 1.43K for all the condition and most scatters are concentrated on [-1,1] for the simulated data. Meanwhile, in situ measurements from Taiyuan Xiaodian and Changwu station were used to evaluate the accuracy of the algorithm under the cloudy-free condition. The RMSE of LST under the cloudy-free condition varies from 3.42 to 5.12 K with a mean errors varying from 0.004 to 3.02K. At the same time, the RMSEs of LST under cloud condition are calculated and the value varies from 7.59 to 7.87 K with the mean errors varying from −3.84 to 3.15 K. Considering the spatial scale inconsistency between the LST from satellite data and the in situ measurements and the sensitivity to instrumental noises, the error can be accepted and the proposed method can be used to retrieve the daytime all-weather LST at high-temporal resolution using geostationary meteorological satellite data without any auxiliary field measurement. The availability of all-weather LST at a high temporal resolution would benefit many research field including climate change, soil moisture and evapotranspiration etc.
Notably, the proposed method assumes that the variation in the LST is caused by variations in insolation (which is related to cloudiness) during the daytime and the parameters of the DTC model of estimated pixel are the same as that of the spatial neighboring-pixel. Therefore, the approach can only be used during the daytime. Furthermore, the DTC parameters should display some differences especially for heterogeneity pixels. How to calibrate the parameters and retrieve the all-weather nighttime LST at high-temporal resolution need to be further deepened and explored.
Funding:
National Natural Science Foundation of China (41271381).
References and links
1. C. Coll, V. García-Santos, R. Niclòs, and V. Caselles, “Test of the MODIS land surface temperature and emissivity separation algorithm with ground measurements over a ricepaddy,” IEEE Trans. Geosci. Remote Sens. 54, 3061–3069 (2016).
2. Z.-L. Li, B.-H. Tang, H. Wu, H. Ren, G. Yan, Z. Wan, I.-F. Trigo, and J.-A. Sobrino, “Satellite-derived land surface temperature: current status and perspectives,” Remote Sens. Environ. 131, 14–37 (2013).
3. S.-B. Duan, Z.-L. Li, and P. Leng, “A framework for the retrieval of all-weather land surface temperature at a high spatial resolution from polar-orbiting infrared and passive microwave data,” Remote Sens. Environ. 195, 107–117 (2017).
4. H. Shwetha and D. Kumar, “Prediction of high spatio-temporal resolution land surface temperature under cloudy conditions using microwave vegetation index and ANN,” ISPRS J. Photogramm. Remote Sens. 117, 40–55 (2016).
5. J.-L. Moncet, P. Liang, J. F. Galantowicz, A. Lipton, G. Uymin, C. Prigent, and C. Grassotti, “Land surface microwave emissivities derived from AMSR-E and MODIS measurements with advanced quality control,” J. Geophys. Res. Atmos. 116, D16104 (2011).
6. M. Neteler, “Estimating Daily Land Surface Temperatures in Mountainous Environments by Reconstructed MODIS LST Data,” Remote Sens. 2, 333–351 (2010).
7. L. Ke, Z. Wang, C. Song, and Z. Lu, “Reconstruction of MODIS LST Time Series and Comparisonwith Land Surface Temperature (T) among Observation Stations in the Northeast Qinghai-Tibet Plateau,” Progress in Geography. 30(7), 819–826 (2011).
8. C. Zeng, H.-F. Shen, M.-L. Zhong, L.-P. Zhang, and P.-H. Wu, “Reconstructing MODIS LST Based on Multitemporal Classification and Robust Regression,” IEEE Geosci. Remote Sens. Lett. 67312, 512–516 (2015).
9. M.-L. Jin, “Interpolation of surface radiative temperature measured from polar-orbiting satellites to a diurnal cycle 2. Cloudy-pixel treatment,” J. Geophys. Res. 105(D3), 4061–4076 (2000).
10. L. Lu, V. Venus, A. Skidmore, T. Wang, and G. Luo, “Estimating land-surface temperature under clouds using MSG/SEVIRI observations,” Int. J. Appl. Earth Obs. 13, 265–276 (2011).
11. W.-P. Yu, M.-G. Ma, X.-F. Wang, and J.-L. Tan, “Estimating the land-surface temperature of pixels covered by clouds in MODIS products,” J. Appl. Earth Obs. 8(1), 083525 (2014).
12. X. Zhang, J. Pang, and L. Li, “Estimation of Land Surface Temperature under Cloudy Skies Using Combined Diurnal Solar Radiation and Surface Temperature Evolution,” Remote Sens. 7(1), 905–921 (2015).
13. Z. Wan and J. Dozier, “A generalized split-window algorithm for retrieving land surface temperature from space. IEEE Trans. Geosci,” Remote Sens. 34, 892–905 (1996).
14. J. A. Sobrino, Z.-L. Li, M. P. Stoll, and F. Becker, “Determination of the surface temperature from ATSR data,” In Proceedings of 25th International Symposium on Remote Sensing of Environment, Graz, Austria, April 4–8. pp. II-19 - II-109(1993).
15. X. Zhang and L. Li, “Estimating net surface shortwave radiation from Chinese geostationary meteorological satellite FengYun-2D (FY-2D) data under clear sky,” Opt. Express 24(6), A476–A487 (2016).
16. S.-B. Duan, Z.-L. Li, N. Wang, H. Wu, B.-H. Tang, X. Jiang, and G. Zhou, “Modeling of day-to-day temporal progression of clear-sky land surface temperature,” IEEE Geosci. Remote Sens. Lett. 10, 1050–1054 (2013).
17. X. Zhang and L. Li, “A method to estimate land surface temperature from Meteosat Second Generation data using multi-temporal data,” Opt. Express 21(26), 31907–31918 (2013).
18. S.-B. Duan, Z.-L. Li, N. Wang, H. Wu, and B.-H. Tang, “Evaluation of six land-surface diurnal temperature cycle models using clear-sky in situ and satellite data,” Remote Sens. Environ. 124, 15–25 (2012).
19. B. Tang, Y. Bi, Z.-L. Li, and J. Xia, “Generalized Split-Window Algorithm for Estimate of Land Surface Temperature from Chinese Geostationary FengYun Meteorological Satellite (FY-2C) Data,” Sensors (Basel) 8(2), 933–951 (2008).
20. Z.-L. Li, L. Jia, Z.-B. Su, Z.-M. Wan, and R.-H. Zhang, “A new approach for retrieving precipitable water from ATSR2 split-window channel data over land,” Int. J. Remote Sens. 24, 5095–5117 (2003).
21. X. Zhang and J. Pang, “A comparison between atmospheric water vapour content retrieval methods using MSG2-SEVIRI thermal IR data,” Int. J. Remote Sens. 36(19–20), 5075–5086 (2015).
