2.1.

Bibliographic Database

For many years, Web of Science (WoS) from the Thomson Reuters Institute for Scientific Information (ISI) was the only citation and publication database covering all domains of science. However, in 2004, Elsevier Science introduced the Scopus database, which is an effective alternative to WoS according to Vieira and Gomes.³⁵ Studies comparing WoS and Scopus have concluded that, when comparing different subject areas, an advantage in one of these databases is shown in Ref. 36. Studies comparing WoS and Scopus which focused on a specific field of study; however, concluded that it is difficult to determine the better database.^37,38 In addition, Vieira and Gomes³⁵ stated that Scopus contains a broader range of journals and provides 20% more coverage than WoS. This finding is consistent with studies by Jacso³⁹ and Falagas et al.,⁴⁰ which confirmed that Scopus covers more journals and analyzes citations faster than WoS. Moreover, Boyle and Sherman⁴¹ recommended that Scopus should be chosen because of its outcome quality, time-saving features, ease of use, and larger amount of search results.

Currently, with almost 23,000 titles and 150,000 books from $\sim 5000$ publishers, Scopus is considered one of the largest abstract and citation databases of peer-reviewed research literature.⁴² We used the Scopus database to select publications for this study (accessed on July 12, 2018).

2.2.

Methods

We followed four crucial steps for a systematic review outlined by Stewart:⁴³ (1) obtain all publications in the interested topic by using broad search keywords, (2) use strict and clear criteria to limit the universal results to targeted/eligible publications, (3) extract and code essential information from the eligible publications for the statistical outcome measures, and (4) report and discuss the methods, results based on the selected publications (Fig. 2).

As stated in Sec. 2.1, in this study, we used the Scopus database to search for publications. It should be noted that using the same search keywords, there are 1101 publications within the timespan 1999 to 2018; however, most of the studies (991 publications or $\sim 90\%$) were published in the period from 2009 to 2018. Therefore, we analyzed publications within this timespan.

Fig. 2

The procedure for selecting articles from Scopus database.

First, we performed an advanced search in the Scopus database on July 12, 2018 using the keywords {“MODIS” AND (“land surface temperature” OR “LST”)} in titles, abstracts, and keywords (including author keywords and index keywords) from 2009 to 2018. This search resulted in 1451 records. Next, we excluded conference papers, book chapters, review articles, letters, notes, and non-English articles, which resulted in 991 peer-reviewed articles (including published articles and articles in press) published in 159 journals. We further filtered this list by removing 136 articles that did not use MODIS LST data. Next, based on the titles, abstracts, and full texts (if necessary), we removed articles that focused more on theories and methods, or only analyzed the spatial and temporal variations of LST. As a result, 529 articles were selected for the final list (Fig. 2).

3.1.

General Research Output Based on Selected Publications

In this section, we present an overview of research application trends using MODIS LST data based on the identified 529 articles published in 159 journals from January 2009 to July 2018.

Generally, the number of publications that used MODIS LST data increased from 2009 to 2017. It should be noted that the publications focusing on theories, methodologies, or validation purposes have been removed from this review study. In 2018, by the date of the Scopus search (July 12), there were 55 articles already published (or in press). It is reasonable to predict that this increasing trend will continue through 2018 (Fig. 3).

Regarding the subject classification of the publications, 529 articles were published within 19 subject areas in 159 journals. The top three subject areas were Earth and Planetary Sciences (397 articles—75%), Environmental Science (172 articles—32.5%), and Agricultural and Biological Sciences (164 articles—31%). The total number of articles was not equal to 529 because many belong to multiple subject areas. In addition, the studies of MODIS LST data were published in both general research area journals (e.g., J-STARS, Hydrology, and Journal of Earth System Science) and specific journals (e.g., Remote Sensing of Environment and International Journal of Climatology), indicating a wide range applications of MODIS LST data.

Table 2 shows the 11 most active journals that have published MODIS LST data-based research. Among the 529 articles, 255 articles (48.2%) were published in these 11 journals, whereas the remaining 274 (51.8%) were published in the other 148 journals. In addition, 65.8% of total citations (TC) were received for these 11 journals, whereas only 34.2% were received for the other 148 journals. The three major publication outlets for MODIS LST data-based research were Remote Sensing of Environment, Remote Sensing, and International Journal of Remote Sensing, with a total number of articles (% articles) of 62 (11.7%), 56 (10.6%), and 30 (5.7%), respectively. Table 2 shows that the journals with the highest TC values were Remote Sensing of Environment with 2808 citations (33.5%), Remote Sensing with 397 citations (4.7%), and the International Journal Of Applied Earth Observation and Geoinformation with 353 citations (4.2%).

Table 2

The top most active journals publishing research based on MODIS LST data.

Table 3 shows the 10 most cited articles, which cover five application topics: UHI effect, Ta estimation, drought monitoring, soil moisture, estimating gross primary production, and net radiation estimation. Among the 10 most cited articles, three articles focus on Ta estimation, two articles focus on UHI effect, and two articles focus on soil moisture applications. Again, it is should be noted that the citation count in this review only used the Scopus database (other databases, e.g., Googlescholar, could show a higher number of citations). As shown in Table 3, the most cited article was written by Imhoff et al.⁴⁴ with 344 citations, followed by Vancutsem et al.⁴⁵ with 199 citations, and Rhee et al.⁴⁶ with 141 citations. These articles studied UHI, Ta estimation, and drought monitoring. An interesting finding is that these top three articles were published in the same year (2010), suggesting that these three topic applications could be the most applications of MODIS LST data.

Table 3

Top 10 most cited articles.

3.2.

Main Applications of MODIS LST Data

In order to get an overview of MODIS LST applications, we generated a word cloud using the titles of the 529 selected articles. As shown in Fig. 4, the most commonly used words were urban, heat, estimation, air, soil, island, and moisture. Based on this result and further investigating the titles, abstracts, and full text (if required) of the 529 selected articles, we found that the top five applications were UHI (63 articles—11.9%), air temperature (Ta) estimation (52 articles—9.8%), soil moisture estimation (50 articles—9.5%), evapotranspiration (48 articles—9.1%), and drought (33 articles—6.2%). This is consistent with the hypothesis we have made based on the citations trend shown in Table 3. The remaining ($\sim 50\%$) were applied for various application fields, such as risk mapping (e.g., earthquakes, forest fires, mosquitoes, and floods); agriculture (rice yields, paddy rice mapping, and rice quality); water (surface water temperature, lake temperature, and water pollution); hydrological models; urban expansion; snow melt; snow cover; and many others.^54–62

Among these top five applications of MODIS LST, three applications could be considered significantly more popular because more than 50% of publications for these applications were published in the last 3 years. These applications are drought monitoring/assessment (21 articles—63.6%), UHI effect (35 articles—55.6%), and Ta estimation (27 articles—52%). However, in comparison with UHI or Ta estimation studies, drought studies used additional data from MODIS sensors (and other sensors). In the other words, MODIS LST is considered a key input for UHI and Ta studies; however, in drought studies, MODIS LST is just one input among many other important inputs (i.e., NDVI from MODIS, precipitation from TRMM).^46,63–66 Therefore, in the next section, we will discuss the UHI effects and Ta estimation/mapping studies.

3.3.

UHI Using MODIS LST Data

The UHI, a well-known effect of urbanization, is a common phenomenon, where surface temperatures are higher in urban areas than in the surrounding rural areas.⁶⁷ This effect is considered one of the most significant impacts of human activity on the Earth’s surface climate.^68,69 Although UHI are localized hotspots in the landscape, they have a profound impact on the lives of the urban residents, who currently account for more than half of the total population; urban population is expected to rise about 66% by 2050.^70,71 In 2014, the number of people living in urban areas in developed regions such as Northern America, Latin America and the Caribbean, and Europe was estimated to be $\sim 82\%$, 80%, and 73%, respectively. Meanwhile, the urban population of developing regions was around 48% and 40% for Asia and Africa, respectively. The UN⁷⁰ further predicted that the fastest urbanization will take place in Asia and Africa, with urban population predictions of about 64% and 54% by 2050, respectively.⁷⁰ This indicates that in these two regions, the urban landscape will change significantly, and therefore resources should be focused there to implement well-planned and sustainable development.

Traditionally, UHI studies were based on the data retrieved from ground weather stations.^72,73 Some limitations of UHI studies using weather station data have been reported in the literature because: (1) the distribution of weather stations, particularly in developing countries or remote regions, is usually sparse; therefore, it cannot represent the UHI for a whole city,⁷⁴ (2) most weather stations are located in urban areas, so rural station data for UHI analysis is lacking,⁷⁵ and (3) quality data from stations are usually not available to all users.⁷³ In contrast, remote sensing satellites provide easy to access LST data with global coverage at no cost and can therefore be considered the most suitable way to investigate the surface UHI (SUHI).

Rao⁷⁶ first reported the perspective of using remote sensing techniques for SUHI studies. In the subsequent years, many scholars stated that, in UHI studies, satellite temperature measurement produced better results than interpolating weather station observations.^77,78 Since the 1990s, the AVHRR sensor has been widely used for LST analysis^79,80 and the SUHI effect at the regional scale.^81,82 Later, ASTER, which provides both LST daytime and nighttime data at high spatial resolution (90 m), has been applied for SUHI studies in many cities.^83–85 Deilami et al.⁸⁶ produced a systematic review of spatiotemporal factors, data, methods, and mitigation measures of the UHI effect, investigating all English peer-reviewed articles published from January 1965 to July 2017. The results showed that the most widely used LST data for Ta were Landsat TM (54.6%), followed by Landsat ETM (34.6%) and MODIS (28.0%). The other satellite images used for UHI studies were Landsat 8,^87–90 SPOT,⁹¹ and QuickBird.⁹² However, as noticed by the USGS in August 2013, surface water temperatures derived from Landsat 8 TIRS data may be inaccurate, potentially impacting the results of research.⁹³ Obviously, this calibration problem has restricted the application of Landsat 8 TIRS for UHI studies.⁹⁴

Although Landsat is considered one of the most popular satellite data sources for UHI studies, according to Deilami et al.,⁸⁶ there are two main shortcomings of Landsat data: (1) the low temporal resolution (16 days) is not suitable for diurnal or weekly UHI monitoring or assessment and (2) due to the small size of Landsat images, multiple images must be mosaiced when applied to regional and national areas. According to Irons et al.⁹⁵ and Wulder et al.,⁹⁶ the mosaic processing takes a significant amount of time. Currently, it is worth noting that the Landsat surface reflectance level-2 products can be retrieved from the NASA website at no cost, but the preprocessing time can be large when requesting multiple images for a large area.

To overcome the limitations of Landsat satellite data over the last decade, MODIS data have been widely applied to SUHI studies, particularly to those focusing on temporal variation, i.e., diurnal, seasonal, annual, decadal, and coarser scales.^50,97–99 In addition, it is worth noting that when investigating the SUHI using remote sensing LST data, not only the surface characteristics should be taken into consideration but also the climatic and meteorological conditions.^100,101 The potential applications and the rising trend of using MODIS data have been reported by Zhang et al.,¹⁰² who conducted a bibliometric analysis of global remote sensing research from 2010 until 2015. To trace the temporal trend of the main applications using remote sensing techniques, they investigated the top keyword from each year (from 2010 to 2015), and the results showed that MODIS emerged as the core theme. The “MODIS” keyword ranked in the top six keywords (among “model,” “classification,” “algorithm,” “SAR,” and “remote sensing”).

As mentioned in Sec. 3.2, one of the most popular applications of MODIS LST data is for UHI/SUHI studies, 12.3% (65 articles). It is worth noting that in UHI/SUHI studies, LST data from Landsat have been widely applied.⁸⁶ However, in recent years, the UHI studies are transitioning from single date to time series (multiple day), from local and regional to global. In addition, the variation of UHI/SUHI due to the changing seasons, daytime and nighttime, has been receiving the attention from scholars. Therefore, remote sensing image data that provide high temporal resolution (i.e., daily, hourly) and coverage of large areas (i.e., city) is becoming indispensable. MODIS satellite data, which provide four LST datasets each day at 1-km spatial resolution with global coverage, are among the most suitable sources for UHI studies at the regional and global scales. Another advantage of MODIS satellite data is that it provides readily available products for UHI applications (e.g., LST and land cover data), which significantly reduces the processing time. Among the 65 articles using MODIS LST data for SUHI studies, 13 articles were focused on analyzing diurnal SUHI, 19 articles analyzed the effects of season on SUHI intensity, and 37 articles investigated the effects of daytime and nighttime on SUHI intensity. Among these 37 articles, two studies (Haashemi et al.⁸ and Bahi et al.¹⁰³) used nighttime LST data of MODIS and daytime LST from Landsat images. It should be noted that among the 65 articles, the numbers of studies that were conducted in China, India, and the United States were 21, 9, and 7, respectively. The remaining articles were mostly implemented in EU cities. However, based on the prediction of UN,⁷⁰ urbanization will take place in developing countries regions, Africa, and Asia; this suggests that MODIS LST will be one of the most valuable data sources for the hot topic, UHI studies, in developing countries in the near future.

Another important point is that there are four MODIS LST data images at four times each day. Between the two satellites (Terra and Aqua) collecting these data, the overpass time of the Aqua satellite (1:30 a.m. and 1:30 a.m.) is closer to the occurrence of the highest and lowest temperatures in the diurnal cycle than the Terra satellite (10:30 a.m. and 10:30 p.m.). However, most studies choose to utilize either Terra data^9,103–107 or Aqua data^{44,73,108–110} without explanation. According to Clinton and Gong,¹¹¹ the daytime SUHI monitored by Aqua LST (1:30 p.m.) was higher than that of Terra (10:30 a.m.). Therefore, a topic, for example, the effects of different time observations on UHI, should be taken into consideration for future studies.

3.4.

Air Surface Temperature Estimation Using MODIS LST Data

Air surface temperature (Ta), which is usually measured at weather stations 2 m above the ground, is a key parameter for a wide range of applications.^48,112–114 Although it provides accurate Ta, it has limitations in spatial coverage. Traditionally, to obtain Ta for a large area, interpolation methods were used;^115,116 however, the accuracy of interpolation methods strongly depends on the distribution of weather stations.^45,117 Therefore, it is difficult to apply this method for regions with sparse weather stations, such as in developing countries and remote areas.

According to Li et al.²⁴ and Chen et al.,¹¹⁸ remotely sensed LST data are considered one of the most suitable ways to retrieve Ta data at local, regional, or global scales. Among the remote sensing satellites, the Advanced Very High Resolution Radiometer (AVHRR), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), and Landsat can provide LST data, though LST data from MODIS are the most widely used.¹⁹ To confirm this, we used the same search terms for ASTER, AVHRR, and Landsat (which resulted in 52 publication using MODIS LST). The result showed that from 2009 to 2018 there were only four records total; three for Landsat and one for ASTER.^119–122 This result indicates that MODIS LST is the most popular and widely used source for Ta estimation studies. Among these four articles, Zhu et al.¹²² used both Landsat 8 and MODIS LST for Ta estimation, showing that Ta could be estimated at the 30-m (using Landsat 8) and 1000-m scales (using MODIS) with a root mean square error (RMSE) of 1.7°C and 2.6°C, respectively. Among the 52 Ta estimation publications collected for this study, the number of studies that applied MODIS LST for daily, weekly, and monthly Ta estimation were 45, 2, and 7, respectively. The total number is not equal to 52 because two studies were applied for both daily and weekly Ta,^48,123 and one study was applied to estimate both daily and monthly Ta.¹²⁴ In addition, the difference between daily, weekly, and monthly articles also suggests that the high temporal resolution of MODIS LST data (four times daily) is one of the largest advantages of MODIS data and is, therefore, widely applied for Ta estimation.

Three common method types were used: statistical methods (78.8%—41 articles), machine learning methods (13.5%—7 articles), and temperature vegetation index methods (7.7%—4 articles). It is worth noting that in most of the articles using statistical or machine learning methods, MODIS LST data were combined with auxiliary variables. The most widely used auxiliary variables were elevation, NDVI, Julian day, latitude, longitude, and day length.^{123,125–127} Although statistical methods were the most widely used, by 2013 the researchers started focusing on machine learning approaches.^127–133 Some studies compared the performance of advanced statistical methods and machine learning methods but with inconclusive results. Xu et al.¹²⁹ compared two methods, linear regression, and random forest (RF) for Ta estimation in British Columbia, Canada, concluding that the RF model achieved better accuracy [mean absolute error (MAE) = 2.02°C, $R^2=074$, compared with MAE = 2.41°C, $R^2=064$]. Meyer et al.¹³⁰ compared a simple linear regression model with three machine learning algorithms [RF, generalized boosted regression models (GBM), and Cubist (CB)] for Ta estimation in Antarctica. Their results showed that machine learning algorithms only slightly outperformed the simple linear estimation model. Among the three machine learning methods, GBM showed the best results for Ta estimation. Zhang et al.¹³¹ compared six statistical models (i.e., multiple linear regression, the partial least squares regression, back propagation neural network, support vector regression, RF, and CB) with two types of MODIS LST data qualities (all clear sky data and only good data). They concluded that the performance of each algorithm depended on the different combinations of MODIS LST and the quality of data. Overall, however, RF and CB showed the best results of Ta estimation with both conditions of data quality. Noi et al.¹³² compared the performance of a linear regression model (LM) with RF and CB algorithms, for two types of data input, MODIS LST data solely and MODIS LST with auxiliary variables (elevation and Julian day). Their results showed that both CB and LM performed well with MODIS LST solely, but when MODIS LST is used with auxiliary variables, CB and RF are recommended. Consistent with Zhang et al.¹³¹ and Noi et al.,¹³² Xu et al.¹³³ compared 10 machine learning algorithms for Ta monthly mapping in the Tibetan Plateau and reported that CB outperformed the other methods.

Regarding the MODIS LST data for Ta estimation, four MODIS LST data are available each day (Terra day, Terra night, Aqua day, and Aqua night), with overpass at local times 10:30 a.m., 10:30 p.m., 1:30 p.m., and 1:30 a.m., respectively. However, most of the studies have chosen only one or two MODIS LST data (among the four available) for Ta estimation. Generally, the selection was based on the overpass time of MODIS LST data. For example, Xu et al.¹²⁹ chose LST of Aqua daytime for Ta-max estimation in British Columbia, Canada; Zhang et al.¹³¹ chose LST of Aqua daytime and Terra nighttime, for Ta-max and Ta-min estimation, respectively; and Yang et al.¹³⁴ chose LST of Aqua daytime and nighttime, for Ta-max and Ta-min estimation, respectively. Recently, some studies have chosen all four MODIS LST data for Ta estimation.^{131,132,135,136}

Most recently, Yoo et al.¹²⁷ proposed a new method for Ta estimation by considering the lag time of heat transport from ground level to 2 m. Therefore, they also used the four LST data taken from the day before, so that they used eight MODIS LST datasets total for daily Ta estimation. They implemented this new method in two megacities, Los Angeles, USA, and Seoul, South Korea, and concluded that LST of the day before were crucial for estimating daily Ta in urban climates.

4.1.

Reconstruction of MODIS LST Data

Regarding the reconstruction techniques, among the 855 articles (included articles focused on theories or methods developments/assessments), there were only 15 articles that considered missing MODIS LST data. Meanwhile, according to Zeng et al.,¹⁴³ completed cloud free MODIS LST data are rare, particularly during rainy seasons and in humid regions. Furthermore, according to Jin and Dickinson,¹⁴⁴ each day, the cloudy-sky condition is present over more than half the globe. This means that MODIS LST data, which have excluded all the cloudy-sky pixels from $M*D11$ products, lose more than half of LST data time series.¹⁴⁵ In addition, $Ø\mathrm{stby}$ et al.¹⁴⁶ compared 8 years of MODIS LST data with in situ observations in Austfonna (an Arctic ice cap located on Svalbard), showing that only 26% of the MODIS LST data are retrieved under actual clear sky conditions, of which the percentages of MODIS LST data under clear sky conditions for winter and summer are about 40% and 20%, respectively. As a result, the subsequent applications that use MODIS LST data will be greatly affected.¹²⁷

Furthermore, many previous studies applied MODIS LST data without any consideration for cloud cover, just simply selecting the measurements under clear sky conditions (using the QC data file) for analysis.^{48,125,131,147}

Among the 15 studies that implemented the reconstruction of MODIS LST data, most only focused on the method development and assessment.^125,148,149 Only a few studies used this technique for specific applications, e.g., Neteler¹⁵⁰ reconstructed MODIS LST for Ta estimation in a mountainous area. In future research, more studies should investigate the effects, and solutions to solve this problem.

4.2.

Downscaling Using MODIS LST Data

Among the 855 articles (including articles focusing on theories and methods), there were 60 articles that applied downscaled methods. Among these, there were three studies on UHI effect,^151–153 11 articles for monitoring/estimation evapotranspiration, and 12 articles for soil moisture applications, whereas the remaining articles mainly focused on theories or method developments/assessments.

Three types of downscaling studies used MODIS LST data: (1) downscaling MODIS LST data to higher spatial resolution (e.g., 250 and 30 m), (2) downscaling MODIS LST data to higher temporal resolution data (e.g., using hourly GOES—LST data), and (3) downscaling other lower spatial resolution products (e.g., SMOS: Soil Moisture and Ocean Salinity, AMSR-E: Advanced Microwave Scanning Radiometer–Earth Observing System).

As LST is considered one of the most crucial parameters for a wide range of applications, at various scales, high spatiotemporal resolution LST data are required. Currently, a number of sensors can provide high spatial resolution (i.e., Landsat, ASTER) or high temporal resolution (e.g., MODIS, GEOS), yet no sensor can provide LST data at both high spatial and temporal resolution. Therefore, downscaling (or fusion) techniques play a crucial role in retrieving high spatial and temporal resolution for many applications. Although Landsat images provide very high spatial resolutions (up to 60 m), the temporal resolution is too coarse (16 days); therefore, MODIS LST are normally downscaled to Landsat resolution.^{16,24–26,154}

Another important point is that in developing countries, the agricultural field size and other land uses are small, and indeed much smaller than the spatial resolution of MODIS LST data. Therefore, one pixel contains many land cover types, receiving a mixed signal. To reduce the effect of this issue, downscaling to finer resolutions is necessary.

4.3.

Quality of MODIS LST Data

According to Colditz et al.,¹⁵⁵ for a long-term time series, data quality is considered a critical issue. MODIS LST data are not available for locations (pixels) that are covered by cloud.¹³⁷ In addition, Wan¹³⁷ also noted that in MODIS LST products, pixels are still present which are covered by thin clouds that cloud-detecting algorithms cannot identify and remove. Therefore, the results of all applications using MODIS LST data are affected. Ackerman et al.¹³⁸ reported that there are $\sim 15\%$ of contamination pixels existent in MODIS LST products. In addition, Williamson et al.¹³⁹ used ground-based meteorological station observations to evaluate the cloud contamination in clear sky MODIS Terra daytime LST in the southwest Yukon area, showing that 13% to 17% of MODIS LST data were unidentified clouds. This will obviously affect the results of subsequent MODIS LST data applications. However, to date, most studies have not considered this problem. Therefore, further studies should investigate the effect on MODIS LST quality and which cloud detection methods are needed.