Creating a seamless 1 km resolution daily land surface temperature dataset for urban and surrounding areas in the conterminous United States

Remote Sensing of Environment 206 (2018) 84–97

Contents lists available at ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Creating a seamless 1 km resolution daily land surface temperature dataset for urban and surrounding areas in the conterminous United States

T

⁎

Xiaoma Lia, Yuyu Zhoua, , Ghassem R. Asrarb, Zhengyuan Zhuc a

Department of Geological and Atmospheric Sciences, Iowa State University, Ames, IA 50011, USA Pacific Northwest National Laboratory, Joint Global Change Research Institute, College Park, MD 20740, USA c Department of Statistics & Statistical Laboratory, Iowa State University, Ames, IA 50011, USA b

A R T I C L E I N F O

A B S T R A C T

Keywords: MODIS Land surface temperature Gapfilling Interpolation Urbanization

High spatiotemporal land surface temperature (LST) datasets are increasingly needed in a variety of fields such as ecology, hydrology, meteorology, epidemiology, and energy systems. Moderate Resolution Imaging Spectroradiometer (MODIS) daily LST is one of such high spatiotemporal datasets that are widely used. But, it has a large amount of missing values primarily because of clouds, shadows, and other atmospheric conditions. Gapfilling the missing values is an important approach to create seamless high spatiotemporal LST datasets. However, current gapfilling methods have limitations in terms of accuracy and efficiency to assemble the data over large areas (e.g., national and continental levels). In this study, we developed a 3-step hybrid method by integrating daily merging (gapfilling missing values at one overpass using values at the other three overpasses each day), spatiotemporal gapfilling (estimating missing values based on values of their spatial and temporal neighbors), and temporal interpolation (gapfilling missing values based on values of their neighboring days), to create a seamless high spatiotemporal LST dataset using the four daily LST observations from the two MODIS instruments on Terra and Aqua satellites. We applied this method in urban and surrounding areas in the conterminous U.S. in 2010. The evaluation of the gapfilled LST product indicates its root mean squared error (RMSE) to be 3.3 K for mid-daytime (1:30 pm) and 2.7 K for mid-nighttime (1:30 am) observations. The method can be easily extended to other years and regions and is also applicable to other satellite products for large areas. This seamless daily (mid-daytime and mid-nighttime) LST product with 1 km spatial resolution is of great value for studying urban climate (e.g., quantifying surface urban heat island intensity, creating seamless high spatiotemporal air temperature dataset) and the related impacts on people (e.g., health and mortality), ecosystems (e.g., phenology), and energy systems (e.g., building energy use).

1. Introduction Land surface temperature (LST), also called radiative skin temperature of ground, is a key indicator of surface energy balance and widely used in a variety of disciplines such as ecology, hydrology, meteorology, epidemiology, and energy systems (Kloog et al., 2014; Li et al., 2017; Peng et al., 2014; Tatem et al., 2004; Zhou et al., 2014b; Zhou et al., 2013; Zhou and Gurney, 2010; Zhou et al., 2012). Solar radiation, atmospheric condition, and land surface characteristics all affect LST, therefore LST has strong spatiotemporal variations (Peng et al., 2014; Zhang et al., 2015). LST products with different spatial and temporal resolutions have been developed from remotely sensed data such as Landsat 8 with 100 m resolution (Roy et al., 2014), Moderate Resolution Imaging Spectroradiometer (MODIS) with 1 km resolution (Wan, 2013), Meteosat Second Generation Spinning Enhanced Visible

⁎

and Infrared Imager (MSG-SEVIRI) with 3 km resolution (Aminou, 2002), and Advance Microwave Scanning Radiometer-Earth Observing System (AMSR-E) with 25 km resolution (Cavalieri et al., 2014). Among them, MODIS LST has some advantages including 1 km spatial resolution, relatively high temporal resolution (four overpasses from two satellites per day), global coverage, and more than ten years of data, which makes it one of the most widely used LST products, especially for large spatial scale (i.e. continental and regional) and long-term (multiseasons to years) studies (Clinton and Gong, 2013; Imhoff et al., 2010; Li et al., 2017; Peng et al., 2011; Tan and Li, 2015; Zhang et al., 2014; Zhao et al., 2014; Zhou et al., 2014a). However, the availability of MODIS LSTs is severely impacted by a number of factors such as clouds, shadows, and other atmospheric conditions, which result in missing LSTs (Crosson et al., 2012; Hu and Brunsell, 2013; Wan, 2013). Temporally aggregating daily LST to a coarser temporal resolution (i.e., 8-

Corresponding author. E-mail address: [email protected] (Y. Zhou).

https://doi.org/10.1016/j.rse.2017.12.010 Received 28 February 2017; Received in revised form 25 October 2017; Accepted 8 December 2017 0034-4257/ © 2017 Elsevier Inc. All rights reserved.

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

efficiency of spatiotemporal gapfilling as it will decrease not only the number of missing data but also the spatial and temporal extent of the gaps. This strategy has been successfully applied to fill the gaps in snow cover (Gao et al., 2010; Parajka and Blöschl, 2008) and create daily air temperature (Ta) dataset (Huang et al., 2015; Zhang et al., 2016), but it has not been integrated with current gapfilling methods to fill MODIS daily LST gap. With the increasing demand for high spatiotemporal LST dataset and the inefficiency of current gapfilling method in filling MODIS daily LST gap for large areas, this study aimed to (1) develop a hybrid gapfilling method, which is more efficient for gapfilling large scale MODIS daily LST; and (2) create a 1 km resolution daily LST dataset at about 1:30 pm (approximating daily maximum) and 1:30 am (approximating daily minimum), local times for urban and surrounding areas in the conterminous U.S. in 2010, but the methodology presented here can be readily applied to other years. The remainder of this paper describes the study area and source data (Section 2), details of the hybrid method (Section 3), results and discussions (Section 4), and conclusions (Section 5).

day, month, seasonal, and yearly) is a widely used approach, to increase spatial coverage of MODIS LST, especially the 8-day LST which is a widely used product. Although this product is advanced for its improved spatial coverage, temporal information could be limited because of using an average temperature. As such, the 8-day LST is not useful in applications that require knowledge of its temporal variations including impacts of temperature on humans, e.g., mortality (Shi et al., 2015), and natural ecosystems, e.g. tree species distribution (Zimmermann et al., 2009). Other studies also suggested that temporally aggregated LSTs may introduce bias, for example, in quantifying the intensity of surface urban heat island (SUHI) (Hu and Brunsell, 2013). Therefore, a seamless dataset of daily LST is still urgently needed, particularly for urban related studies. A number of methods and algorithms have been developed and applied to fill the gaps in remotely sensed data (Gerber et al., 2016; Poggio et al., 2012; Shen et al., 2015; Weiss et al., 2014). They can generally be grouped into four categories. The first category combines data of different overpasses of the same satellite or data from different satellites on the same day (Crosson et al., 2012; Duan et al., 2017; Kou et al., 2016; Shwetha and Kumar, 2016). The second category estimates the missing values based on empirical relationship between available LST and auxiliary data, such as latitude, longitude, elevation, normalized difference vegetation index (NDVI), and surface moisture (Fan et al., 2014; Ke et al., 2013). The third category fills the gaps by relying on temporal, spatial, or spatiotemporal information of the LST dataset itself using temporal interpolation (Kilibarda et al., 2014; Xu and Shen, 2013), spatial interpolation (Ke et al., 2013; Yang et al., 2004), and spatiotemporal gapfilling methods (Sun et al., 2017; Weiss et al., 2014; Zeng et al., 2015). The fourth category is hybrid method(s), which combines several methods mentioned above (Metz et al., 2014; Neteler, 2010; Weiss et al., 2014). Among these, the hybrid method is most promising for filling MODIS daily LST gap as it utilizes the advantages of these methods but avoids their disadvantages. Currently, high spatiotemporal (e.g., 1 km and daily) LST datasets are not widely available. One main reason is the challenge of effectively filling MODIS LST gap, while ensuring a high quality data product, especially for large areas. MODIS daily LST has significant spatiotemporal variations, including long term trend (e.g., seasonal trend), and short term variations (e.g., daily and diurnal variation). Most gapfilling methods currently are designed for reconstructing the missing value of datasets (parameters) with relatively small or no short term variation (e.g., vegetation index (Gao et al., 2008; Julien and Sobrino, 2010), snow cover (Gao et al., 2010; Parajka and Blöschl, 2008)), therefore they are not effective for reconstructing daily LST which has high temporal dynamics. For example, temporal interpolation can completely fill the gaps rapidly, but it has difficulty in reconstructing the fine scale temporal variation and may miss high or low LSTs (Metz et al., 2014; Xu and Shen, 2013). Spatiotemporal gapfilling appears to be the most suitable method to reconstruct the missing value of remotely sensed data with high spatiotemporal variation (Sun et al., 2017; Weiss et al., 2014). It has been applied to fill gaps in Landsat ETM + SLC-off images (Chen et al., 2011; Zhu et al., 2012), but it requires a significant amount of computing time for filling MODIS daily LST gaps, because of their large spatial and long temporal extent (Weiss et al., 2014). Monthly aggregated LST was used to decrease the computing time cost by Weiss et al. (2014) for gapfilling 8-day LST in Africa. However, temporally aggregating LST will lead to losing of fine scale variation and may introduce extra uncertainty. MODIS LST from Terra and Aqua satellites, with four daily observations (i.e. four overpasses), provides an opportunity to fill the missing values for a given time using values from the other three times. For example, Crosson et al. (2012) used Terra LST to fill Aqua LST gap and increased data coverage of Aqua LST by 24% and 30% for daytime and nighttime overpasses, respectively, in the conterminous U.S. Therefore, adding an additional step of merging four daily observations of MODIS LST provides a potential opportunity to increase the

2. Study area and data We focused our analysis of LSTs on urban and surrounding areas in the conterminous U.S. Urban areas are vulnerable to environmental changes and influenced strongly by weather/climate conditions. For example, temperature in the urban areas is usually higher than that in the rural areas, known as urban heat island (UHI) effect. The dynamics of urban systems in turn affect their environment, and those of surrounding areas (i.e. sub-urban systems), for example through the formation of SUHI. However, we have limited understanding of the spatiotemporal characteristics and driving factors of SUHI and their environmental impacts due to a lack of required high spatiotemporal resolution temperature datasets in urban systems. In this study, we focused on developing a seamless 1 km resolution daily LST dataset, which will help us better understand high spatiotemporal characteristics of the urban thermal environment. The focus on urban areas will help test the effectiveness of the proposed gapfilling method because of larger gaps in such areas, as compared to the rural areas (Hu and Brunsell, 2013; Shen et al., 2015). In addition, the focus on conterminous U.S., instead of a single region or city, will help us test the robustness of the proposed method because of a large number of urban areas with different degrees of heterogeneity. We also included the surrounding sub-urban areas to be able to assess the effects of urbanization on the urban environmental conditions (e.g., SUHI). The urban areas in this study were defined by the urban map developed based on the Defense Meteorological Satellite Program/ Operational Linescan System (DMSP/OLS), nighttime stable light data (NTL) in 2010 using a cluster method (Zhou et al., 2014c; Zhou et al., 2015). The surrounding sub-urban areas were defined to have an extent similar to the urban area, but beyond its boundary, i.e. a buffer zone with equal size to its corresponding urban area (Li et al., 2017; Peng et al., 2011; Zhou et al., 2014a). The source dataset is MODIS 1 km resolution daily LST product V6 from instruments on the National Aeronautics and Space Administration (NASA) Earth Observing System (EOS) Terra and Aqua satellites (MOD11A1 and MYD11A1). The MODIS LST is derived from two thermal infrared bands 31 (10.78–11.28 μm) and 32 (11.77–12.27 μm) using the generalized split-window algorithm (Wan, 2013, 2014; Wan and Dozier, 1996). These daily LST data include four overpasses with two daytime imagery (i.e. 10:30 am on Terra (T1) and 1:30 pm on Aqua (T2)) and two nighttime imagery (i.e. 10:30 pm on Terra (T3) and 1:30 am on Aqua (T4)), all in local crossing times. The LST data together with ancillary information (i.e., view angle and view time) were obtained from Land Processes Distributed Active Archive Center (LP DAAC). 85

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

MYD11A1

Pre-procession

(Terra & Aqua)

View angle mask

View time normalization T1 (Terra Day)

T2 (Aqua Day)

T3 (Terra Night)

Daily merging

Shift

Shift

Regression

T4 (Aqua Night)

Daily merged T2

1

Spatiotemporal gapfilling

Spatiotemporal gapfilling

Spatiotemporal gapfilled T2

Yes Is there gap? Yes

Is gapfilling

criterion satisfied?

No

No

Temporal interpolation

Temporal interpolation

Completed T2

Fig. 1. Schematic illustration of the hybrid gapfilling method, using T2 as an example. (This schematic illustration is also applicable for T4 using the regression method for T3 and the shift method for T1 and T2).

3. Methods

3.1. Step 1: data pre-processing

Considering data availability and the advantage and disadvantage of different gapfilling methods (or algorithms), we developed a hybrid method to build the seamless daily LST. This hybrid method includes four steps: (1) pre-processing: minimizing the impacts of view angle and view time; (2) daily merging: merging or compositing four MODIS daily LSTs; (3) spatiotemporal gapfilling: filling the gaps based on the spatiotemporal information of the daily merged LST; and (4) temporal interpolation: filling the remaining gaps using a temporal interpolation method based on the output of spatiotemporal gapfilling. Fig. 1 illustrates the steps for this hybrid method by showing how a gap in T2 is filled, as an example. In the following section, we described details for each step.

The first step is pre-processing the MODIS LST source data to minimize the effects of view angles and times (Duan et al., 2014a; Hu et al., 2014). MODIS daily LST data has view angles ranging from 0 to 65 degrees, which has significant impacts on the retrieved LST (Hu et al., 2014; Jin, 2004). We only used LST with view angle smaller than 45 degrees, because view angle effect on LST is relatively smaller for angles smaller than 45 degrees (Jin, 2004). Another source of uncertainty in MODIS LST data is caused by variation of viewing time. For example, view time for T1 is assumed as 10:30 am local time, however, it usually varies from 10:00 am to 12:00 pm local time (Duan et al., 2014a). Two methods can be used to temporally normalize daily LSTs at different view times: diurnal temperature change (DTC) model (Duan et al., 2014b; Jin and Dickinson, 1999), and linear regression model (Duan et al., 2014a). The limited number of daily overpasses and 86

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

days of each month. We estimated daily shift by temporally interpolating monthly averaged shift to include variations of LSTs. For example, to estimate T2 using T3, we first calculated daily LST difference (shift) between T2 and T3 (black point in the central subplot in Fig. 2). Then, we estimated the average shift for each month and linearly interpolated them to estimate the shift for days on which T2 is not available (green line in the central subplot in Fig. 2). Finally, the estimated shift was added to T3 as the estimation of T2 using T3. After applying the linear regression and shift method, we obtained four time series of T2 (the third column in Fig. 2), and composited them in the priority order of observed T2, and estimated T2 from T1, T3, and T4 as the final LSTs by daily merging (the fourth column in Fig. 2). Similar approach was applied to gapfill missing values of T4 using a linear regression method for T3 and a shift method for T1 and T2. The four T4 LSTs were composited in the order of observed T4, and estimated T4 from T3, T2, and T1.

relatively large number of missing LST values prohibit the use of DTC model. We applied the linear regression model and used the parameters (slopes) obtained by Duan et al. (2014a) to normalize T1 to 11:00 am local time, assuming a linear change of LST from 10:00 am to 12:00 pm local time. The study of Duan et al. (2014a) indicated that this method can achieve a root mean squared error (RMSE) of 0.5 K in estimating LST at 11:00 am local time. We did not temporally normalize T2 because of its nonlinear change during the period of view times as shown in DTC curves (Duan et al., 2014b; Jin and Dickinson, 1999), and the absence of needed correction parameters. We did not correct the effect of view time on T3 and T4, because they are small, which is verified by DTC curves (Duan et al., 2014b; Jin and Dickinson, 1999). A quality assessment (QA) flag is usually used to select high quality LST data. We did not filter LSTs based on QA because the QA in urban areas is demonstrated to be only partially reliable (Bechtel, 2017), and there are very few available LSTs if, for example, a QA threshold of LST error < 1 K is applied.

3.3. Step 3: spatiotemporal gapfilling 3.2. Step 2: daily merging We completed the spatiotemporal gapfilling based on the spatial and temporal information from the daily merged LSTs. We chose spatiotemporal gapfilling instead of spatial interpolation or temporal interpolation because of its higher accuracy. Two powerful spatiotemporal gapfilling algorithms, the algorithm from R Gapfill package (Gerber et al., 2016) and the python coded algorithm (Weiss et al., 2014) are available for this purpose. We used the R Gapfill package for two major reasons: (1) the comparison by Gerber et al. (2016) shows that the algorithm from the Gapfill package performs better, and (2) the python coded algorithm uses monthly averaged data as a reference, which may lose temporal variations, and no temporal aggregation needed in the Gapfill package. The Gapfill package is designed to fill the missing values in remote sensing data based on observations within a spatial and temporal window expressed in four-dimensions (x (longitude), y (latitude), day, and year). The gapfilling process is performed for each pixel and the procedure includes two steps; (1) sub-setting the

Two methods can be used to estimate LST at a given time based on the values at three other times: (1) DTC model (Duan et al., 2014b; Jin and Dickinson, 1999), and (2) linear regression (Coops et al., 2007) or shift (Crosson et al., 2012). Considering data availability and computing efficiency, we used linear regression (Coops et al., 2007) and shift (Crosson et al., 2012) method. Fig. 2 shows the daily merging for T2 by applying this procedure to one pixel as an example. To estimate T2 based on T1, we first built a linear regression model with T2 as dependent variable and T1 as independent variable based on available time series of LSTs in a year, then this relationship was used to estimate the missed T2 if T1 is available. Due to the non-linear relationship between daytime and nighttime LSTs (i.e. T2 and T4), we applied the shift method (Crosson et al., 2012) to estimate the missed T2 based on available T3 and T4. Crosson et al. (2012) used seasonally averaged shift to estimate daily LSTs, which means an equal shift is used for all

Fig. 2. Daily merging of four LST observations to gapfill T2, applying to a single pixel as an example. LST observation of four overpasses (First column), estimation model (linear regression and shift) with LST of another three overpasses (second column), estimated LST of the four overpasses (third column), and the final estimation (fourth column). A similar approach is applied for T4 with a regression method for T3 and a shift method for T1 and T2.

87

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Fig. 3. Schematic illustration of the spatiotemporal gapfilling method, using one pixel of missing value marked as a “red star” on day 226 as an example. Sub-setting of the data is based on a window size of λ1 = λ2 = 10, λ3 = 4, and λ4 = 0 (A), for this example. The procedure for quantile regression (B), is adapted from Gerber et al. (2016). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

be higher than λ5, (2) number of available LST for the area/image under consideration must be higher than λ6. If both criteria are satisfied, the quantile regression is performed and the missing value is calculated. Otherwise, the spatial window size (λ1 = λ2) increases until both criteria are fulfilled. However, increasing the spatial window size will also increase required computing time and may also result in decreased accuracy in the calculated LST. To increase gapfilling efficiency, another parameter, λ7 is set to define the maximum window radius increase, beyond which the gapfilling process is considered unsuccessful. Considering the high spatiotemporal variation of LST, according to published literatures (Gerber et al., 2016; Zeng et al., 2015) and sensitivity analysis for the spatial and temporal window radius (λ3 and λ7, see details in Supplementary data), we set these parameters to: (1) λ1 = λ2 = 10 pixels, (2) λ3 = 4, which means that each subset has 9 potential areas/days, (3) λ4 = 0, which indicates only LST of the studied year is used as input considering the spatial pattern of LST in urban and surrounding areas in other years could be greatly different, (4) λ5 = 4, which means the 9 areas/days of the subset must have at least 4 areas/days with LST data, (5) λ6 = 5, which means the area/ days under consideration must have at least 5 LST observations, and (6) λ7 = 20, which constraints the maximum spatial window size to be 30 (λ1 + λ7) pixels in all directions. There exist some outliers (speckles) in the spatiotemporally gapfilled LSTs. We identified and removed these outliers using a time series

data using a four-dimensional window defined by λ1, λ2, λ3, and λ4, which correspond to the dimensions of x, y, day, and year in all directions, respectively. Fig. 3A shows an example of the sub-setting for filling the gap marked as a red star on day 226 with λ1 = λ2 = 10, λ3 = 4, and λ4 = 0; (2) estimating the missing value using a quantile regression method based on the selected subset of data. Fig. 3B shows an example for this procedure. Score values were first calculated for each of the nine subsets through pixel-wise comparisons of non-missing values, assuming the values of each subset have similar spatial but shifted temporal patterns. Based on these score values, the subsets are ranked, and the rank values are recorded. Then the quantile at the location of the missing value (i.e., the red star), is calculated for each subset and averaged. A quantile regression model is then built using the observed LSTs as a dependent variable and the ranks of the subsets as an explanatory variable, based on the estimated quantile. Finally, this regression is applied to estimate the LST value for the gap. The details of the algorithm can be found in Gerber et al. (2016). The spatiotemporal window size determines the performance of the gapfilling method, including percent gaps filled, accuracy, and computing time cost (Gerber et al., 2016). Gapfill adopts a dynamic spatial window size, which adapts to the local distribution of missing values. An initial window size is first defined by λ1, λ2, λ3, and λ4, and then two criteria are evaluated: (1) number of non-empty area/image, with minimum requirement of having at least one available LST value, must 88

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

merging can also reduce the spatial and temporal extent of the gaps, therefore, it can increase the efficiency of spatiotemporal gapfilling, in a subsequent step. This will be discussed in the following paragraphs. Spatiotemporal gapfilling has an equal role to the daily merging, accounting for 36% of the study area. Finally, temporal interpolation fills the remaining 1% gaps (Fig. 4). The number of observed and gapfilled LSTs shows significant spatial variation. Taking T2 on day 204 as an example, observed LSTs cover 37.2% of the entire U.S. There are large number of missing values in many states in Western (e.g., California, Washington), Midwest (e.g., Missouri, Arkansas), and Eastern (e.g., New Jersey, Massachusetts) regions of U.S. (Fig. 5A). By merging daily LSTs from the four overpasses, LST coverage increases to 65% for the entire U.S. Daily merging fills most of the missing values in Western and Midwest U.S. but fails in filling the gaps in Eastern region because no LSTs are available for all the four satellite overpasses (Fig. 5B). Fortunately, the spatiotemporal gapfilling completely fills the remaining gaps for all regions using available LSTs in their spatial and temporal neighbors (Fig. 5C). The spatial distribution of observed and gapfilled LSTs for T4 on day 204 is similar to that for T2 (Fig. S5 in Supplementary data). The LST availability also displays temporal variations. Seasonally, summer and fall have > 30% observations, while data availability in winter is < 20% (Fig. 6). Similar temporal pattern is observed for pixels filled by daily merging, i.e., summer and fall have the largest percent of filled data, followed by spring and winter. Spatiotemporal gapfilling plays a larger role in winter, filling > 50% of the total missing pixels. There are more variations in data availability on a daily basis (Fig. S6 in Supplementary data) than that at seasonal level. In addition to the number of missing LSTs, the temporal extent (consecutive days) of missing LSTs also shows strong spatial and temporal variations. Taking three urban sites in Atlanta, Los Angeles and New York City (NYC) as examples, all three sites, for both T2 and T4, have missing LSTs for > 200 days, with maximum temporal extent ranging from 10 days for T4 in Los Angeles to 24 days for T4 in NYC (Fig. 7, Table 1). Daily merging increases LST coverage by > 100 days (107–152), and LST coverage increases to about 200 days in NYC and to > 240 days in the other two cities. Daily merging fills almost all missing values for the summer period in Los Angeles. However, there is still a large number of missing LST values especially in Atlanta and NYC (> 100 days). Fortunately, the temporal extent of the gap decreases significantly to < 10 days, which enhances the efficiency of the following spatiotemporal gapfilling. The subsequent spatiotemporal gapfilling fills all remaining missing values, and no additional temporal interpolation is needed (Table 1).

analysis, instead of spatial analysis, because the latter approach may introduce bias in the LST variations due to factors such as land cover types, topography, and contaminated pixels (Hu et al., 2014). In the time series analysis, we first fitted a third order polynomial curve for the observed T2/T4 to remove the temporal (seasonal) trend, and calculated standard deviation of the residual (observed residual). Then, we calculated the residual of the gapfilled LST (gapfilled residual) based on the previously fitted third order polynomial curve. Pixels with the gapfilled residual outside 3 standard deviations of the observed residual were treated as outliers and removed. Three standard deviations were determined by considering the balance between omission and commission of the outliers. More details of the outlier determination are listed in Supplementary data. We tested two different approaches to fill the remaining gaps because the spatiotemporal gapfilling algorithm could not fill all the gaps, mainly because there were not sufficient number of available LSTs within the 30 pixels radius window. One approach was to repeat the spatiotemporal gapfilling by using the gapfilled LSTs as input (Sun et al., 2017; Weiss et al., 2014), and the second approach was to fill the gaps using temporal interpolation (the third step). We found the former have higher accuracy with RMSE of 3.5 K (T2) and 3.3 K (T4) by repeating spatiotemporal gapfilling, lower than 5.1 K (T2) and 4.5 K (T4) by temporal interpolation. Therefore, we repeated the spatiotemporal gapfilling with the filled LST as input until no extra gaps could be filled. 3.4. Step 4: temporal interpolation After the spatiotemporal gapfilling, a small percentage of gaps (about 1% of the study area) still exist, especially in areas where gaps are large and the urban areas (patches) are small and scattered. Increase of the spatial window size in the spatiotemporal gapfilling can gapfill all missing LSTs, but it is not efficient for a small percentage of gaps as it requires a large amount of computing time. Therefore, we used a temporal interpolation method to fill the remaining gaps considering its better efficiency for a small percentage of gaps. 4. Results and discussion 4.1. Gapfilling daily LST using the hybrid method Different steps of the hybrid method have different roles in filling the LST gaps. Available MODIS daily LST covers < 30% (28.4% and 28.3% for T2 and T4, respectively) of the study area (Fig. 4). This number is lower than that summarized by Crosson et al. (2012) for the conterminous U.S. for two reasons: (1) data availability is intrinsically lower in urban areas than that in rural areas (Hu and Brunsell, 2013; Xu and Shen, 2013), and (2) we did not include LSTs with view angle larger than 45 degrees. Daily merging of the four observations increases LST coverage by 123.1% and 124.7% for T2 and T4, respectively, which cover 35% and 35.2% of the study area, correspondingly (Fig. 4). Daily

A: Percent pixel for T2

4.2. Validation and uncertainty of the gapfilled LSTs We conducted a stepwise accuracy assessment for the gapfilled LSTs to estimate the combined uncertainty caused by the algorithms themselves, and error propagation when the output of previous steps is used

0.94 11.23

0.94 12.14

28.43

24.40 4.92

Fig. 4. Percent of total number of pixels (i.e. area) for the observed and gapfilled LSTs in 2010, for the conterminous U.S., based on proposed approach in this study (STP1: daily merging, STP2: spatiotemporal gapfilling, STP3: temporal interpolation, STP2_S1: STP2 using output of STP1 as input, STP2_S2: repeating STP2 using the output of STP2 as input).

B: Percent pixel for T4 28.25

23.45 14.14

19.36 10.72

13.05

Observed

STP1_T1

STP1_T3

STP2_S1

STP2_S2

STP3

STP1_T4

8.03

Observed

STP1_T3

STP1_T2

STP2_S1

STP2_S2

STP3

89

STP1_T1

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Percent (%)

Fig. 5. Spatial distribution of observed and gapfilled LSTs (T2) on day 204.

100

1.0

80

35.3

0.5

0.9

28.5

27.6

Finally, we obtained higher RMSE values as we focused on the urban areas, where LSTs have more variation and uncertainty than that in the rural areas. The RMSE values for our method are, however, comparable to the daily estimation using spatially coarser LST retrieved from passive microwave data (Kou et al., 2016; Shwetha and Kumar, 2016). There are large variations of the RMSE among different steps. The day-day merge and night-night merge show the lowest RMSE (2.2 K for T2 and 1.5 K for T4) (Table 2). This is not surprising as the viewing times of the overpasses are close (about 3 h) (Crosson et al., 2012). However, RMSEs for day-night merge is much larger and is estimated to be > 4 K, for example when estimating T2 by T4 (Table 2). There are two possible reasons for this: (1) there is high probability of change in environmental conditions (e.g., cloud and wind) between the daytime and nighttime LSTs, which increases the uncertainty of estimating daytime LST using nighttime LST and vice versa; (2) a temporal interpolation is used to estimate the daily day-night shift/difference, which may generate larger uncertainty because only temporal information is used. Theoretically, this shift/difference can be estimated using the spatiotemporal gapfilling method, however, the large number of missing values prohibits this approach. The spatiotemporal gapfilling (step 2) results in high accuracy as expected, demonstrating the importance of integrating both spatial and temporal information simultaneously to fill the missing values in spatiotemporal data (Sun et al., 2017; Weiss et al., 2014). The third step, temporal interpolation shows the largest uncertainty because of the simplicity of the temporal interpolation algorithm itself, and the error propagation. Together with the gapfilled LSTs, we also produced an associated layer which records the corresponding LSTs (i.e., T1, T2, T3, or T4) used by daily merging and gapfilling steps (i.e., first round spatiotemporal gapfilling, repeated spatiotemporal gapfilling, and temporal interpolation). Therefore, the accuracy of gapfilled LSTs at the pixel level can be inferred based on the value of the associated layer and the corresponding RMSE.

1.4

51.3 60 37.9 40

38.3

35.2 28.6

20 28.0

33.4

33.7 17.8

0 Spring Observed

Summer

Spatiotemporal gapfilling

Fall Daily merging

Winter

Temporal interpolation

Fig. 6. Seasonal variations of LST availability based on the proposed hybrid method in this study for year 2010 (Spring: March, April, May; Summer: June, July, August; Fall: September, October, November; Winter: January, February, December).

as input of the following steps. Taking T2 as an example, we assumed all observations of T2 were missing and performed the hybrid gapfilling using the output as input. The finally estimated T2 is compared with the observed T2 to estimate the accuracy of each step for 20 randomly selected days. The overall RMSEs are 3.3 K for T2 and 2.7 K for T4, with the mean absolute error (MAE) of 2.4 K for T2 and 1.8 K for T4, with corresponding R2 of 0.93 for T2 and 0.94 for T4, respectively (Fig. 8). The RMSE values are higher than those gapfilling MODIS 8-day LSTs (Weiss et al., 2014; Xu and Shen, 2013), mostly due to a larger number of missing data and larger spatiotemporal extent of daily LST gaps (Weiss et al., 2014). In addition, the daily LST has much higher variation than the 8-day LST, which also increases the calculated RMSE. 90

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Fig. 7. Temporal pattern of observed and gapfilled LSTs: in Atlanta (first row), Los Angeles (second row) and New York City (third row); (OBS: observed-red, STP1: daily merging-green; STP2: spatiotemporal gapfilling-blue dots). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

added artificial gaps in LSTs for 20 randomly selected days using gaps of another 20 random days, then filled these gaps using the three methods, respectively. Finally RMSE, percent gaps filled, and computing time were estimated and compared for the three methods, taking T2 of MODIS tile “h11v04” as an example (Table 3). Temporal interpolation shows the least computing time cost, but the largest RMSE (about 4.0 K). This is understandable for the simplicity of the algorithm, which only uses information from neighboring days. Spatiotemporal gapfilling shows similar RMSE (4.0 K) to temporal interpolation. This is because there are large gaps in the input LST and the spatiotemporal gapfilling needs to repeat the gapfilling routine many times to fill these gaps. For example, in 17 of 20 tested days it repeats > 10 times. If we only perform the spatiotemporal gapfilling once using the observed LST as input, we can obtain a low RMSE (1.9 K), but only 40% of gaps can be filled. Accordingly, the efficiency of the spatiotemporal gapfilling decreases with the increase of repeats because of the large size of gaps. The proposed hybrid method effectively solves these problems by merging daily LSTs of four observations, which not only fills the gaps directly, but also decreases the size of the gaps (Table 1). After merging daily LSTs, the average number of repeats in the spatiotemporal gapfilling decreases from 22 to 4 for the tested days, resulting in a higher

The accuracy of the gapfilled LSTs was further evaluated by comparing the temporal variation between the gapfilled LST and Ta from Global Historical Climatology Network-Daily (GHCND) observations (Neteler, 2010; Shwetha and Kumar, 2016). Taking three selected weather stations in Atlanta, Los Angeles, and NYC as examples, the gapfilled LSTs display similar seasonal trend and daily variations as compared with Ta (Fig. 9). The magnitude of LST and Ta values are different, and LST usually shows larger range and variation than Ta. For example, T2 is larger than maximum Ta, and T4 is smaller than minimum Ta. This is understandable because LST and Ta are driven by different mechanisms (Gallo et al., 2011; Good, 2016). It should be noted that variations of LST and Ta do not match well in some days. This is probably due to different impacts of factors such as land cover and environmental conditions on temporal characteristics of LST and Ta (Gallo et al., 2011; Good, 2016; Oyler et al., 2016). 4.3. Comparing with other gapfilling methods We compared the performance (e.g., RMSE, percent gaps filled, and computing time) of the hybrid method against two other widely used methods: temporal interpolation and spatiotemporal gapfilling. We

Table 1 Total number of LST gaps, maximum gap, and filled LSTs by each step for the three example urban sites. Observed (days)

Atlanta Los Angeles New York City

T1 T4 T1 T4 T1 T4

Daily merging (days)

Spatiotemporal gapfilling (days)

Gaps

Max gap

Filled

Gaps

Max gap

Filled

Gaps

259 265 232 217 272 290

13 21 11 10 17 24

138 142 152 136 107 122

121 123 80 81 165 168

7 7 9 9 9 9

121 123 80 81 165 168

0 0 0 0 0 0

91

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Fig. 8. A comparison of observed and gapfilled daily LSTs for T2 and T4.

Table 2 Stepwise accuracy assessment of the gapfilled LST (step1: daily merging, step2: spatiotemporal gapfilling, step3: temporal interpolation, S1: spatiotemporal gapfilling using daily merged LST as input; S2: repeating spatiotemporal gapfilling using the output as input). T2

All steps Daily merging T1 T3 T4 Spatiotemporal gapfilling S1 S2 Temporal interpolation

Table 3 A comparative analysis of the performance of three gapfilling methods (Taking T2 of MODIS tile h11v04 as an example). Method

RMSE (K)

Percent gaps filled (%)a

Computing time costb

Temporal interpolation Spatiotemporal gapfilling The hybrid method

3.99 3.95

100 77

0.00008 4.48

3.35

100

1

T4 RMSE (K)

R2 (%)

RMSE (K)

R2 (%)

3.29 3.32 2.15 3.79 4.14 3.08 3.01 3.51 5.35

93.31 92.86 97.75 88.17 85.78 94.52 94.72 93.39 93.11

2.68 2.51 1.5 3.72 3.21 3.36 3.38 3.29 4.06

93.5 93.92 97.92 86.2 86.54 92.07 92.87 87.98 83.31

T3 T2 T1 S1 S2

a b

Pixels of filled LST including observed LST divided by total pixel in the study area. Relative to the computing time cost by the hybrid method.

accuracy and less computing time, as compared to the other two methods. We performed independent assessment of accuracy for each step of the hybrid method to better understand sources of uncertainties. We

Fig. 9. Temporal patterns of Ta and gapfilled LSTs: in Atlanta (first row), Los Angeles (second row), and New York City (third row).

92

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

the spatial distribution of land areas covered by forests and croplands in the Eastern region mitigating the effects of solar heating. Third, urban areas are hotter than their surrounding areas, which is consistent with the widely observed and reported influence of UHI phenomenon. The seasonal mean gapfilled LSTs show a similar spatial pattern to that of the annual mean gapfilled LSTs (Figs. S7–S8 in Supplementary data). In addition to the regional variation, the gapfilled LSTs also display strong local variation as observed in three example metropolitans illustrated in Fig. 11 for T2 on day 204. The consistently higher LSTs are distributed in urban centers and decrease with the increase of distance to the center of urban areas. In addition, the LST map displays heat archipelagos with strong variation even in the highly urbanized areas. It should be noted that LSTs in these three regions are from different sources: Atlanta is from observation, Los Angeles from daily merging, and NYC from spatiotemporal gapfilling. This demonstrates different steps of the hybrid gapfilling method can reconstruct the missing LSTs very effectively. One noticeable feature of the gapfilled LST dataset based on our proposed method as compared with other seamless temperature datasets is its ability to capture the details of spatial heterogeneity in urban areas. Since there is no other seamless and daily LST dataset available for this purpose, we compared the derived gapfilled LSTs with Daymet, a 1 km resolution daily Ta dataset (Thornton et al., 2016). The gapfilled LSTs show a stronger spatial variation than the Daymet (Fig. 11). This is probably because Ta is intrinsically less heterogenous than LST, and the Daymet data is produced based solely on elevation and does not take into account the changes in land surface characteristics (e.g., urbanization impact on land use and land cover). It has been reported that LST can be a valuable source data for creating high spatiotemporal Ta dataset (Good, 2015; Kilibarda et al., 2014). Given that most currently available LST datasets are either too coarse in resolution (spatial or temporal) or they often have too many missing values, the resulting LST dataset in this study could fill this void and serve as a very good sources for estimating spatial detail of Ta, especially in the urban areas, which is usually ignored in most current gridded Ta datasets such as Daymet. Another advantage of the resulting gapfilled LST dataset from this study is the ability to capture the impact of local extreme events which is missing in the current and widely used 8-day LST datasets (Wan et al., 2015). Fig. 12 shows the temporal pattern of the gapfilled LST from this study as compared with the 8-day LST, for the three urban sites in Atlanta, Los Angeles, and New York City, as an example. Both datasets capture the seasonal trend of LST being higher in summer as compared to winter. However, the gapfilled LST captures the daily variation for extreme events, which are critical for urban environment modeling and management. It should be noted that the 8-day LST still has missing values, for example in Atlanta (Fig. 12), while the gapfilled LST completely fills such gaps.

Table 4 Independent accuracy assessment of each step of the hybrid method. T2

Daily merging

Spatiotemporal gapfilling Temporal interpolation

T1 T3 T4

T4 RMSE (K)

R2

1.47 2.44 2.72 2.00 2.45

0.96 0.93 0.93 0.93 0.95

T3 T2 T1

RMSE (K)

R2

1.44 2.90 2.24 1.53 2.27

0.99 0.96 0.98 0.98 0.98

added artificial gaps in LSTs in 20 randomly selected days using gaps in another 20 random days and performed daily merging, spatial temporal gapfilling, and temporal interpolation using the same LST input, respectively. We calculated RMSE between observed and gapfilled LSTs for pixels that can be filled using these three methods. Similar to the result of stepwise accuracy assessment, this independent accuracy assessment shows that the day-day merging and night-night merging have the highest accuracy, followed by spatiotemporal gapfilling (Table 4). However, we found that the day-night merging shows similar RMSE to that of temporal interpolation (Table 4). One possible reason is that the daily day-night LST shift/difference is also estimated using a temporal interpolation method. We further tested the hybrid method without day-night merging and found that: (1) day-night merging can both increase and decrease the overall accuracy, depending on the accuracy of day-night merging; that is, including day-night merging can increase the accuracy of the hybrid method if day-night merging can achieve high accuracy; (2) this approach can reduce computing time significantly for the later spatiotemporal gapfilling step because it decreases the number of gaps and also the spatial and temporal extents of gaps. Therefore, we included the day-night merging even though it may increase errors in some pixels, considering that it offers significant improvement in efficiency in later spatiotemporal gapfilling step especially for large scale applications (e.g., the conterminous U.S.), and given that the overall accuracy is comparable with previous studies (Kou et al., 2016; Shwetha and Kumar, 2016). A promising research opportunity in future is to improve the overall efficiency with reduced uncertainties in day-night merging approach. 4.4. Spatiotemporal patterns of gapfilled LSTs and their comparisons with independent datasets The annual mean gapfilled LSTs for conterminous U.S. in 2010 displays a clear spatial pattern (Fig. 10). First, it follows latitudinal gradient with lower values in higher latitude areas because of the change of solar radiation. Second, LSTs in Eastern U.S. are lower than those in Western U.S., especially in daytime, which is consistent with

Fig. 10. Spatial pattern of annual mean gapfilled LSTs for the conterminous U.S. in 2010.

93

Fig. 11. Examples of the gapfilled LSTs (T2) and maximum Ta based on Daymet data in three U.S. metropolitans (Atlanta, Los Angeles, and Philladelphia) on day 204, in year 2010.

X. Li et al.

Remote Sensing of Environment 206 (2018) 84–97

94

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Fig. 12. Temporal patterns of the gapfilled daily LST as compared with MODIS 8-day LST: in Atlanta (first row), Los Angeles (second row), and New York City (third row).

5. Conclusions

efficiently. Its application in the urban and surrounding areas in the conterminous U.S. shows high efficiency with satisfactory accuracy, however, it has limitations that can be further improved in future. First, estimating daytime LST based on nighttime LST and vice versa can increase computing efficiency significantly, but could also be a major source of errors in the hybrid method (Table 2). Better algorithm(s) can be developed to improve overall efficiency and accuracy, particularly for day-night merging with consideration of using additional environmental datasets such as cloud cover and wind. Second, the spatiotemporal gapfilling is the most time-consuming among the algorithms currently used. Improving the efficiency of the spatiotemporal gapfilling algorithm, or developing new algorithms could be a worthwhile avenue of research (Sun et al., 2017; Weiss et al., 2014). It should be noted that similar to other LST gapfilling methods (Crosson et al., 2012; Metz et al., 2014; Sun et al., 2017; Weiss et al., 2014; Xu and Shen, 2013; Zeng et al., 2015), the gapfilled LSTs mostly represent those under clear sky conditions and may overestimate LSTs for cloudy sky conditions (Metz et al., 2014). Developing high spatiotemporal LSTs under cloudy sky conditions continues to be a major challenge, but, the gapfilled clear sky LSTs fill an existing research gap for studies of urban areas. Despite its limitations, the hybrid method produces a dataset with a high probability of capturing extreme conditions, which have significant impacts on people, built environment, and the ecosystem, especially in urban areas. Other propducts, such as LSTs retrieved from Passive Microwave brightness temperature measurements, such as those from AMSR-E (Duan et al., 2017; Kou et al., 2016; Shwetha and Kumar, 2016), can be integrated with our gapfilled LSTs to better represent LSTs under cloudy sky conditions.

In this study we developed a 3-step hybrid method, including daily merging, spatiotemporal gapfilling, and temporal interpolation, to fully fill the gaps in MODIS daily LST based on four daily LSTs observations from NASA EOS Terra and Aqua satellites. We applied this method to urban and surrounding areas of the conterminous U.S. for year 2010, and produced a high spatiotemporal (1 km and daily) LST dataset representing daily maximum and minimum corresponding to mid-daytime and mid-nighttime LSTs, respectively. We found the method robustly fills the gaps in existing LST datasets, especially for relatively large areas. It fills all the missing values which cover > 70% of the study area, with a RMSE of 3.3 K, MAE of 2.4 K, and overall R2 of 0.93 for mid-daytime (1:30 pm), and RMSE of 2.7 K, MAE of 1.8 K, and overall R2 of 0.94 for mid-nighttime (1:30 am). The proposed hybrid method can be easily extended to other years and other study areas, and it is also applicable for other remotely sensed data with high spatiotemporal resolution/variation. The proposed hybrid method has four major advantages compared with current methods used for gapfilling MODIS daily LSTs. First, the spatiotemporal gapfilling, which integrates both spatial and temporal information, results in higher accuracy compared to other methods that use either spatial or temporal information only. Second, the proposed method performs very well for daily LST, and captures daily LST variation and extreme events, better than methods based on temporally aggregated LST (e.g., 8-day or monthly averaged LST). Third, a daily merging step is included to decrease computation cost of the spatiotemporal gapfilling step, and makes the method more efficient especially for large areas. Finally, the proposed method uses only LST as input and it does not rely on other auxiliary datasets, reducing the difficulty of acquiring such datasets that have their own inherent uncertainties, thus uncertainty of the final dataset. The proposed hybrid method aims to integrate multiple complementary algorithms to fill MODIS daily LST for large areas more

Acknowledgements This study was supported by the NASA ROSES LULC Program “NNH11ZDA001N-LCLUC” and the U.S. Department of Energy, Office of Science, Biological and Environmental Research as part of the 95

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Integrated Assessment Research and Earth System Modeling programs. We thank Florian Gerber for the help in using the R gapfill package. We would like thank three anonymous reviewers for their constructive comments and suggestions, and the many colleagues and organizations that shared the data used in this project. The views and opinions expressed in this paper are those of the authors alone.

Qinghai-Tibet Plateau using geostatistical approach. IEEE Geosci. Remote Sens. Lett. 10, 1602–1606. Kilibarda, M., Hengl, T., Heuvelink, G.B.M., Gräler, B., Pebesma, E., Perčec Tadić, M., Bajat, B., 2014. Spatio-temporal interpolation of daily temperatures for global land areas at 1 km resolution. J. Geophys. Res. Atmos. 119, 2294–2313. Kloog, I., Nordio, F., Coull, B.A., Schwartz, J., 2014. Predicting spatiotemporal mean air temperature using MODIS satellite surface temperature measurements across the Northeastern USA. Remote Sens. Environ. 150, 132–139. Kou, X., Jiang, L., Bo, Y., Yan, S., Chai, L., 2016. Estimation of land surface temperature through blending MODIS and AMSR-E data with the Bayesian maximum entropy method. Remote Sens. 8, 105. Li, X., Zhou, Y., Asrar, G.R., Mao, J., Li, X., Li, W., 2017. Response of vegetation phenology to urbanization in the conterminous United States. Glob. Chang. Biol. 23, 2818–2830. Li, X., Zhou, Y., Asrar, G.R., Imhoff, M., Li, X., 2017. The surface urban heat island response to urban expansion: a panel analysis for the conterminous United States. Sci. Total Environ. 605, 426–435. Metz, M., Rocchini, D., Neteler, M., 2014. Surface temperatures at the continental scale: tracking changes with remote sensing at unprecedented detail. Remote Sens. 6, 3822–3840. Neteler, M., 2010. Estimating daily land surface temperatures in mountainous environments by reconstructed MODIS LST data. Remote Sens. 2, 333–351. Oyler, J.W., Dobrowski, S.Z., Holden, Z.A., Running, S.W., 2016. Remotely sensed land skin temperature as a spatial predictor of air temperature across the conterminous United States. J. Appl. Meteorol. Climatol. 55, 1441–1445. Parajka, J., Blöschl, G., 2008. Spatio-temporal combination of MODIS images – potential for snow cover mapping. Water Resour. Res. 44, W03406. Peng, S., Piao, S., Ciais, P., Friedlingstein, P., Ottle, C., Bréon, F.-M., Nan, H., Zhou, L., Myneni, R.B., 2011. Surface urban heat island across 419 global big cities. Environ. Sci. Technol. 46, 696–703. Peng, S.-S., Piao, S., Zeng, Z., Ciais, P., Zhou, L., Li, L.Z.X., Myneni, R.B., Yin, Y., Zeng, H., 2014. Afforestation in China cools local land surface temperature. Proc. Natl. Acad. Sci. 111, 2915–2919. Poggio, L., Gimona, A., Brown, I., 2012. Spatio-temporal MODIS EVI gap filling under cloud cover: an example in Scotland. ISPRS J. Photogramm. Remote Sens. 72, 56–72. Roy, D.P., Wulder, M.A., Loveland, T.R., W., C.E., Allen, R.G., Anderson, M.C., Helder, D., Irons, J.R., Johnson, D.M., Kennedy, R., Scambos, T.A., Schaaf, C.B., Schott, J.R., Sheng, Y., Vermote, E.F., Belward, A.S., Bindschadler, R., Cohen, W.B., Gao, F., Hipple, J.D., Hostert, P., Huntington, J., Justice, C.O., Kilic, A., Kovalskyy, V., Lee, Z.P., Lymburner, L., Masek, J.G., McCorkel, J., Shuai, Y., Trezza, R., Vogelmann, J., Wynne, R.H., Zhu, Z., 2014. Landsat-8: science and product vision for terrestrial global change research. Remote Sens. Environ. 145, 154–172. Shen, H., Li, X., Cheng, Q., Zeng, C., Yang, G., Li, H., Zhang, L., 2015. Missing information reconstruction of remote sensing data: a technical review. IEEE Geosci. Remote Sens. Mag. 3, 61–85. Shi, L., Kloog, I., Zanobetti, A., Liu, P., Schwartz, J.D., 2015. Impacts of temperature and its variability on mortality in New England. Nat. Clim. Chang. 5, 988–991. Shwetha, H.R., Kumar, D.N., 2016. 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. Sun, L., Chen, Z., Gao, F., Anderson, M., Song, L., Wang, L., Hu, B., Yang, Y., 2017. Reconstructing daily clear-sky land surface temperature for cloudy regions from MODIS data. Comput. Geosci. 105, 10–20. Tan, M., Li, X., 2015. Quantifying the effects of settlement size on urban heat islands in fairly uniform geographic areas. Habitat Int. 49, 100–106. Tatem, A.J., Goetz, S.J., Hay, S.I., 2004. Terra and Aqua: new data for epidemiology and public health. Int. J. Appl. Earth Obs. Geoinf. 6, 33–46. Thornton, P.E., Thornton, M.M., Mayer, B.W., Wei, Y., Devarakonda, R., Vose, R.S., Cook, R.B., 2016. Daymet: daily surface weather data on a 1-km grid for North America, version 3. In: ORNL Distributed Active Archive Center. Wan, Z., 2013. Collection-6 MODIS Land Surface Temperature Products Users' Guide. Earth Research Institute, University of California, Santa Barbara, CA. Wan, Z., 2014. New refinements and validation of the collection-6 MODIS land-surface temperature/emissivity product. Remote Sens. Environ. 140, 36–45. Wan, Z., Dozier, J., 1996. A generalized split-window algorithm for retrieving land-surface temperature from space. IEEE Trans. Geosci. Remote Sens. 34, 892–905. Wan, Z., Hook, S., Hulley, G., 2015. MYD11A2 MODIS/Aqua Land Surface Temperature/ Emissivity 8-day L3 Global 1 km SIN Grid V006. NASA EOSDIS Land Processes DAAC. http://dx.doi.org/10.5067/MODIS/MYD11A2.006. Weiss, D.J., Atkinson, P.M., Bhatt, S., Mappin, B., Hay, S.I., Gething, P.W., 2014. An effective approach for gap-filling continental scale remotely sensed time-series. ISPRS J. Photogramm. Remote Sens. 98, 106–118. Xu, Y., Shen, Y., 2013. Reconstruction of the land surface temperature time series using harmonic analysis. Comput. Geosci. 61, 126–132. Yang, J., Wang, Y., August, P., 2004. Estimation of land surface temperature using spatial interpolation and satellite-derived surface emissivity. J. Environ. Inf. 4, 37–44. Zeng, C., Shen, H., Zhong, M., Zhang, L., Wu, P., 2015. Reconstructing MODIS lst based on multitemporal classification and robust regression. IEEE Geosci. Remote Sens. Lett. 12, 512–516. Zhang, P., Bounoua, L., Imhoff, M.L., Wolfe, R.E., Thome, K., 2014. Comparison of MODIS land surface temperature and air temperature over the continental USA meteorological stations. Can. J. Remote. Sens. 40, 110–122. Zhang, X., Pang, J., Li, L., 2015. Estimation of land surface temperature under cloudy skies using combined diurnal solar radiation and surface temperature evolution. Remote Sens. 7, 905–921. Zhang, H., Zhang, F., Ye, M., Che, T., Zhang, G., 2016. Estimating daily air temperatures

Appendix A. Supplementary Material Supplementary data to this article can be found online at https:// doi.org/10.1016/j.rse.2017.12.010. References Aminou, D., 2002. MSG's SEVIRI Instrument. ESA Bulletin (0376-4265). pp. 15–17. Bechtel, B., 2017. Data of the climatology of the MODIS land surface temperature version v0.7. http://icdc.cen.uni-hamburg.de/1/daten/land/modis-lst-annualcycle.html. Cavalieri, D.J., Markus, T., Comiso, J.C., 2014. AMSR-E/Aqua Daily L3 25 km Brightness Temperature & Sea Ice Concentration Polar Grids. Version 3. [Indicate Subset Used]. NASA National Snow and Ice Data Center Distributed Active Archive Center, Boulder, Colorado USA. http://dx.doi.org/10.5067/AMSR-E/AE_SI25.003. Chen, J., Zhu, X., Vogelmann, J.E., Gao, F., Jin, S., 2011. A simple and effective method for filling gaps in Landsat ETM + SLC-off images. Remote Sens. Environ. 115, 1053–1064. Clinton, N., Gong, P., 2013. MODIS detected surface urban heat islands and sinks: global locations and controls. Remote Sens. Environ. 134, 294–304. Coops, N.C., Duro, D.C., Wulder, M.A., Han, T., 2007. Estimating afternoon MODIS land surface temperatures (LST) based on morning MODIS overpass, location and elevation information. Int. J. Remote Sens. 28, 2391–2396. Crosson, W.L., Al-Hamdan, M.Z., Hemmings, S.N.J., Wade, G.M., 2012. A daily merged MODIS Aqua–Terra land surface temperature data set for the conterminous United States. Remote Sens. Environ. 119, 315–324. Duan, S.-B., Li, Z.-L., Tang, B.-H., Wu, H., Tang, R., 2014a. Generation of a time-consistent land surface temperature product from MODIS data. Remote Sens. Environ. 140, 339–349. Duan, S.-B., Li, Z.-L., Tang, B.-H., Wu, H., Tang, R., Bi, Y., Zhou, G., 2014b. Estimation of diurnal cycle of land surface temperature at high temporal and spatial resolution from clear-sky MODIS data. Remote Sens. 6, 3247–3262. Duan, S.-B., Li, Z.-L., Leng, P., 2017. A framework for the retrieval of all-weather land surface temperature at a high spatial resolution from polar-orbiting thermal infrared and passive microwave data. Remote Sens. Environ. 195, 107–117. Fan, X.-M., Liu, H.-G., Liu, G.-H., Li, S.-B., 2014. Reconstruction of MODIS land-surface temperature in a flat terrain and fragmented landscape. Int. J. Remote Sens. 35, 7857–7877. Gallo, K., Hale, R., Tarpley, D., Yu, Y., 2011. Evaluation of the relationship between air and land surface temperature under clear- and cloudy-sky conditions. J. Appl. Meteorol. Climatol. 50, 767–775. Gao, F., Morisette, J.T., Wolfe, R.E., Ederer, G., Pedelty, J., Masuoka, E., Myneni, R., Tan, B., Nightingale, J., 2008. An algorithm to produce temporally and spatially continuous MODIS-LAI time series. IEEE Geosci. Remote Sens. Lett. 5, 60–64. Gao, Y., Xie, H., Yao, T., Xue, C., 2010. Integrated assessment on multi-temporal and multi-sensor combinations for reducing cloud obscuration of MODIS snow cover products of the Pacific Northwest USA. Remote Sens. Environ. 114, 1662–1675. Gerber, F., Furrer, R., Schaepman-Strub, G., de Jong, R., Schaepman, M.E., 2016. Predicting Missing Values in Spatio-temporal Satellite Data. arXiv preprint arXiv:1605.01038. Good, E., 2015. Daily minimum and maximum surface air temperatures from geostationary satellite data. J. Geophys. Res. Atmos. 120, 2306–2324. Good, E.J., 2016. An in situ-based analysis of the relationship between land surface “skin” and screen-level air temperatures. J. Geophys. Res. Atmos. 121 (2016JD025318). Hu, L., Brunsell, N.A., 2013. The impact of temporal aggregation of land surface temperature data for surface urban heat island (SUHI) monitoring. Remote Sens. Environ. 134, 162–174. Hu, L., Brunsell, N.A., Monaghan, A.J., Barlage, M., Wilhelmi, O.V., 2014. How can we use MODIS land surface temperature to validate long-term urban model simulations? J. Geophys. Res. Atmos. 119, 3185–3201. Huang, R., Zhang, C., Huang, J., Zhu, D., Wang, L., Liu, J., 2015. Mapping of daily mean air temperature in agricultural regions using daytime and nighttime land surface temperatures derived from TERRA and AQUA MODIS data. Remote Sens. 7, 8728–8756. Imhoff, M.L., Zhang, P., Wolfe, R.E., Bounoua, L., 2010. Remote sensing of the urban heat island effect across biomes in the continental USA. Remote Sens. Environ. 114, 504–513. Jin, M., 2004. Analysis of land skin temperature using AVHRR observations. Bull. Am. Meteorol. Soc. 85, 587. Jin, M., Dickinson, R.E., 1999. Interpolation of surface radiative temperature measured from polar orbiting satellites to a diurnal cycle: 1. Without clouds. J. Geophys. Res. Atmos. 104, 2105–2116. Julien, Y., Sobrino, J.A., 2010. Comparison of cloud-reconstruction methods for time series of composite NDVI data. Remote Sens. Environ. 114, 618–625. Ke, L., Ding, X., Song, C., 2013. Reconstruction of time-series MODIS LST in central

96

Remote Sensing of Environment 206 (2018) 84–97

X. Li et al.

Zhou, Y., Clarke, L., Eom, J., Kyle, P., Patel, P., Kim, S.H., Dirks, J., Jensen, E., Liu, Y., Rice, J., 2014b. Modeling the effect of climate change on US state-level buildings energy demands in an integrated assessment framework. Appl. Energy 113, 1077–1088. Zhou, Y., Smith, S.J., Elvidge, C.D., Zhao, K., Thomson, A., Imhoff, M., 2014c. A clusterbased method to map urban area from DMSP/OLS nightlights. Remote Sens. Environ. 147, 173–185. Zhou, Y., Smith, S.J., Zhao, K., Imhoff, M., Thomson, A., Bond-Lamberty, B., Asrar, G.R., Zhang, X., He, C., Elvidge, C.D., 2015. A global map of urban extent from nightlights. Environ. Res. Lett. 10, 054011. Zhu, X., Liu, D., Chen, J., 2012. A new geostatistical approach for filling gaps in Landsat ETM + SLC-off images. Remote Sens. Environ. 124, 49–60. Zimmermann, N.E., Yoccoz, N.G., Edwards, T.C., Meier, E.S., Thuiller, W., Guisan, A., Schmatz, D.R., Pearman, P.B., 2009. Climatic extremes improve predictions of spatial patterns of tree species. Proc. Natl. Acad. Sci. 106, 19723–19728.

over the Tibetan Plateau by dynamically integrating MODIS LST data. J. Geophys. Res. Atmos. 121 (2016JD025154). Zhao, L., Lee, X., Smith, R.B., Oleson, K., 2014. Strong contributions of local background climate to urban heat islands. Nature 511, 216–219. Zhou, Y., Gurney, K., 2010. A new methodology for quantifying on-site residential and commercial fossil fuel CO2 emissions at the building spatial scale and hourly time scale. Cardiol. Manage. 1, 45–56. Zhou, Y., Weng, Q., Gurney, K.R., Shuai, Y., Hu, X., 2012. Estimation of the relationship between remotely sensed anthropogenic heat discharge and building energy use. ISPRS J. Photogramm. Remote Sens. 67, 65–72. Zhou, Y., Eom, J., Clarke, L., 2013. The effect of global climate change, population distribution, and climate mitigation on building energy use in the U.S. and China. Clim. Chang. 119, 979–992. Zhou, D., Zhao, S., Liu, S., Zhang, L., Zhu, C., 2014a. Surface urban heat island in China's 32 major cities: spatial patterns and drivers. Remote Sens. Environ. 152, 51–61.

97