1. Introduction
Land surface temperature (LST) is one of the key parameters in land–surface physical processes on regional and global scales and has been widely applied to hydrology, meteorology, and the surface energy balance [
1,
2,
3]. Remote sensing is a unique way of obtaining the LST at regional and global scales. Various LST products produced from different satellite data have been widely used in the urban ecological environment, water management, and natural disasters [
4,
5,
6,
7].
Landsat 8 (formerly the Landsat Data Continuity Mission, LDCM) was launched in 2013. Combined with other Landsat series, it provides continuity with the more than 40-year-long Landsat land imaging data set [
8]. The thermal infrared sensor (TIRS) with two thermal infrared channels was added to the Landsat 8 payload to support the detection of the urban heat island, volcanoes, and forest fires. Researchers have developed many algorithms to retrieve the LST from Landsat 8 data, for example, the single-channel algorithm [
9,
10,
11,
12,
13], the split-window algorithm [
12,
14,
15], and the temperature and emissivity separation method [
16]. Meanwhile, some verification work is also underway. According to a study by Meng et al. [
9], the average bias and root mean square error (RMSE) of the estimated LST derived by the radiative transfer equation (RTE) method are 0.09 K and 2.20 K for band 10, respectively. The study by Cook et al. [
17] indicated that the RTE-estimated LST under cloud-free conditions has a bias (standard deviation) of −0.56 K (0.76 K) for band 10 and −2.16 K (1.64 K) for band 11. The study by Parastatidis et al. [
18] indicated the bias and RMSE were 0.1 K and 1.31 K, respectively, for Landsat 8 LST retrieved using the single-channel algorithm (SCA) proposed by Jiménez-Muñoz et al. [
19]. Wang et al. [
20] found that the RMSEs lie between 1.7 K and 4.7 K and 3.3 K and 8.9 K for LST retrieved by the RTE and SCA from band 10, respectively.
Unfortunately, stray light from far out-of-field has affected the absolute calibration of the Landsat 8 TIRS since its launch. Barsi et al. [
21] found a large error in two TIRS bands, −2.1 K (−4.4 K) at 300 K in band 10 (band 11). Vicarious calibration of Landsat 8 TIRS bands by Sobrino et al. [
22] indicated that a bias (RMSE) of 0.01 (0.12) and −0.16 (0.25)
existed in bands 10 and 11, respectively. Li et al. [
23] used the Infrared Atmospheric Sounding Interferometer (IASI) /Metop-B hyperspectral channels to intercalibrate against two Landsat 8 TIRS bands, and the intercalibration biases of the brightness temperature are −0.54
1.21 K and −1.52
1.35 K for bands 10 and 11, respectively. Much effort has been taken to develop an algorithm to remove this stray light. For example, Montanaro et al. [
24] determined the cause of stray light artifacts in 2014, and Gerace et al. [
25] developed an algorithm that was implemented into the processing system in 2017. Through stray light correction, errors were reduced from 2.1 K to 0.3 K at 300 K for band 10 and from 4.4 K to 0.19 K for band 11 (
https://landsat.usgs.gov/april-25-2017-tirs-stray-light-correction-implemented-collection-1-processing). From then on, scholars have carried out some work on the verification of those developed algorithms using corrected data. According to the validation of Duan et al. [
26], the biases (RMSEs) of RTE-derived LST are approximately −0.2 (1.2) and −0.5 K (1.0 K) at the 1 km pixel scale over sand and grassland, respectively. García-Santos et al. [
27] have compared the accuracy of RTE, SCA, and the split-window algorithm proposed by Jiménez-Muñoz et al. [
12] (JMS) and Du et al. [
14] (Du). The results indicated the biases (RMSEs) of LST are −1.4 (2.0) and 0.4 K (1.6 K) for Du and JMS, respectively. The estimated LSTs have a bias and RMSE of −0.1 (2.0) and 2.3 K (3.6 K) when the RTE method was applied to bands 10 and 11, respectively. When SCA applied to TIRS band 10 proposed by Jiménez-Muñoz et al. [
12] and Wang et al. [
13], the biases and RMSEs are 0.8 (0.7) and 2.2 K (2.3 K).
Unlike the single-channel algorithm, the split-window algorithm does not require high-precision atmospheric profiles [
28], and it has been demonstrated that using the split-window algorithm for atmospheric correction can minimize errors in LST retrieval [
15,
29]. Moreover, as pointed out by Jiménez-Muñoz et al. [
12], the split-window algorithm performs well over global conditions and with better results than the single-channel algorithm in theory. Therefore, the split-window algorithm was selected in this paper. Although some split-window algorithms have been proposed [
12,
14,
15], which algorithm is more suitable for Landsat 8 LST estimates remains unknown. Therefore, there are still two problems that need to be resolved. First, the validations were almost always done though simulated datasets, lacking validation with ground-measured data [
12,
14,
15]. Second, the general accuracy of split-window algorithms in real applications remains unclear. Therefore, comprehensive investigations are needed to validate the accuracy of Landsat 8 LST derived from the split-window algorithm.
Recently, the National Oceanic and Atmospheric Administration (NOAA) Joint Polar Satellite System (JPSS) Land Environmental Data Records (EDR) team is developing an enterprise LST algorithm that will be used for both the JPSS and Geostationary Operational Environmental Satellite-R (GOES-R) satellite missions [
30,
31]. This provides an alternative method to retrieve land surface temperature from Landsat 8 TIRS data. Meanwhile, more field campaigns have been conducted to collect the flux measurements, which makes the comprehensive validation of LST possible. This study aims to adapt the enterprise algorithm to the estimate of Landsat 8 LST and evaluate its performance use in situ measurements. The structure of this paper is organized as follows. We will describe the enterprise algorithm, land surface emissivity estimation, and ground LST estimation in
Section 2. The results and analysis are presented in
Section 3. The discussion and conclusions are presented in
Section 4 and
Section 5.
4. Discussion
In this section, we compare the enterprise algorithm with the results of other studies. For example, Zhang et al. [
66] validated Landsat 8 LSTs retrieved from a single-channel algorithm proposed by Jiménez-Muñoz et al. [
12] (hereafter JMS) using SURFRAD sites. The bias, mean absolute error (MAE), and RMSE of the JMS method are 1.49, 1.57, and 1.96 K, respectively. We selected the same 21 Landsat-8 images in SURFRAD as Zhang et al. [
66], and the validation results indicated the Enterprise algorithm, Wan and Sobrino have similar performance. The biases (RMSEs) were 0.85 (2.34), 0.74 (2.28), and 0.94 K (2.46 K) for the Enterprise algorithm, Wan, and Sobrino, respectively.
Yu et al. [
67] compared three different approaches for LST inversion, including the radiative transfer equation algorithm applied to band 10 (hereafter RTE10) and band 11 (hereafter RTE11), the split-window algorithm proposed by Qin et al. [
68] (hereafter Qin) and the single-channel method proposed by Jiménez-Muñoz et al. [
19] applied to band 10 (hereafter JMS10) and band 11 (hereafter JMS11). The LST derived by RTE applied to band 10 in our result is called RTE_M, hereafter. In RTE_M, the MERRA2-6 reanalysis product was used for atmospheric correction. We have selected 20 images, all of which exist in our work and in the study of Yu et al. The biases (RMSEs) between the retrieved LSTs using various methods and in-situ LSTs are shown in
Table 6. The results show that the LSTs retrieved from the RTE method (RTE10, RTE11, and RTE_M) and the single-channel algorithm (JMS10 and JMS11) have the higher accuracy with a bias lower than 1K, while four split-window algorithms (Qin, Wan, the Enterprise algorithm, and Sobrino) have low precision with bias (RMSE) that varies between −1.28 K (2.93 K) and 2.08 K (3.09 K). The accuracy of the Enterprise algorithm, Wan, and Sobrino are similar, and the biases (RMSEs) are 1.90 K (2.93 K), 1.85 K (2.95 K) and 2.08 K (3.09 K), respectively.
García-Santos et al. [
27] also compared three methods for estimating the LST from Landsat-8 data, including RTE10, RTE11, the single-channel algorithm proposed by Jiménez-Muñoz et al. [
19] (hereafter JMS) and Wang et al. [
13] (hereafter Wang) and the split-window algorithm proposed by Jiménez-Muñoz et al. [
12] (hereafter JM2014), and Du et al. [
14] (hereafter Du). The result of García-Santos and our result using 21 Landsat-8 scenes in Mallorca Island is shown in
Table 7. The accuracy of RTE_M was nearly the same as that of RTE10, and the bias and RMSE were 0.1 (−0.1) and 2.3 K (2.3 K). The method of Du had higher absolute bias (RMSE) than the Enterprise algorithm, Wan, and Sobrino and the differences were 1.2 and 0.3 K for bias and RMSE. In this validation, the Enterprise algorithm, Wan, and Sobrino achieved similar bias and equal RMSE. Three split-window algorithms and JM2014 achieved a similar accuracy, with a bias and RMSE of 0.2 (0.4) and 1.7 K (1.6 K), respectively. From existing verification results, we can conclude that the precision of the Enterprise algorithm, Wan, and Sobrino had similar accuracy in most cases.
Although the validation accuracy was relatively high, some phenomena in Landsat-8 images are worth mentioning. The striping and banding were still very noticeable, and the LST differences on two sides of the strip were not negligible. In addition, as noted by Gerace and Montanaro [
25], even the implementation of the stray light algorithm has improved the performance of the TIRS instrument, and the TIRS sensor will continue to be monitored to ensure the expected radiation accuracy for all users.
In spite of single-channel algorithms having better performance in finite verification results, we recommend split-window algorithms for LST retrieval. The reasons are as follows: As pointed out by Jiménez-Muñoz et al. [
12], the split-window algorithm has better results than the single-channel algorithm in theory. Then, Sobrino and Jiménez-Muñoz [
69], Jiménez-Muñoz and Sobrino [
70], and Li et al. [
71] found that all of the single-channel algorithms provide poor results at high atmospheric water vapor contents, and the split-window algorithm performs well over global conditions [
12]. In addition, most of the in-situ sites have low water vapor contents, and more ground data needs to be collected for comprehensive analysis, especially in high water vapor content areas.
5. Conclusions
The NOAA JPSS Enterprise algorithm is adapted to Landsat-8 data to obtain the estimate of LST in this study. The improved normalized difference vegetation index-based threshold method was used to estimate the LSE of non-vegetated areas, and the vegetation cover method was adopted to calculate the LSE of vegetated areas. The coefficient of the enterprise algorithm was derived from the simulation dataset using GAPRI atmospheric profiles. Both the simulation dataset and ground measurements were used to test and validate the enterprise algorithm. In addition, the enterprise algorithm was compared to the generalized split-window algorithm and the split-window algorithm of Sobrino.
Simulated datasets derived from three independent atmospheric profiles (CLAR, TIGR, and SeeBor) were used to test the performance of the enterprise algorithm. The biases (RMSEs) of the Enterprise algorithm were between −0.16 (0.42) and 0.30 K (1.42 K), and the biases (RMSEs) of the generalized split-window algorithm were all between −0.31 (0.36) and 0.43 K (1.21 K). For the method of Sobrino, the biases (RMSEs) range from −0.26 (0.48) to 0.33 K (1.18 K).
The in-situ LSTs derived from the flux measurements at SURFRAD, HiWATER sites, and BanGe site were used to validate the Landsat-8 LSTs retrieved by the Enterprise algorithm and other algorithms. The biases (RMSEs) of the Enterprise algorithm were 1.38 (3.22), 1.01 (2.32), 1.99 (3.49), 2.53 (3.46), and −0.15 K (1.11 K) at the SURFRAD, HiWATER_A, HiWATER_B, HiWATER_C sites, and BanGe site, respectively. For the generalized split-window algorithm, the biases (RMSEs) were 1.39 (3.20), 1.0 (2.30), 1.93 (3.48), 2.53 (3.35), and −0.35 K (1.16 K), respectively, whereas those values were 1.45 (3.39), 1.08 (2.41), 2.16 (3.67), 2.52 (3.58), and 0.02 K (1.12 K) for the split-window algorithm of Sobrino.
Both the test and validation results show that the Enterprise algorithm have similar performance with the other two split-window algorithms. Regarding the existing verification results, we can conclude that the Enterprise algorithm can be used to retrieve Landsat-8 LSTs with a similar accuracy to the generalized split-window algorithm and the split-window algorithm of Sobrino. This study provides an alternative method to estimate LSTs from Landsat-8 data.