Estimation of Long-Term Surface Downward Longwave Radiation over the Global Land from 2000 to 2018

Abstract

It is of great importance for climate change studies to construct a worldwide, long-term surface downward longwave radiation (L_d, 4–100 μm) dataset. Although a number of global L_d datasets are available, their low accuracies and coarse spatial resolutions limit their applications. This study generated a daily L_d dataset with a 5-km spatial resolution over the global land surface from 2000 to 2018 using atmospheric parameters, which include 2-m air temperature (Ta), relative humidity (RH) at 1000 hPa, total column water vapor (TCWV), surface downward shortwave radiation (S_d), and elevation, based on the gradient boosting regression tree (GBRT) method. The generated L_d dataset was evaluated using ground measurements collected from AmeriFlux, AsiaFlux, baseline surface radiation network (BSRN), surface radiation budget network (SURFRAD), and FLUXNET networks. The validation results showed that the root mean square error (RMSE), mean bias error (MBE), and correlation coefficient (R) values of the generated daily L_d dataset were 17.78 W m⁻², 0.99 W m⁻², and 0.96 (p < 0.01). Comparisons with other global land surface radiation products indicated that the generated L_d dataset performed better than the clouds and earth’s radiant energy system synoptic (CERES-SYN) edition 4.1 dataset and ERA5 reanalysis product at the selected sites. In addition, the analysis of the spatiotemporal characteristics for the generated L_d dataset showed an increasing trend of 1.8 W m⁻² per decade (p < 0.01) from 2003 to 2018, which was closely related to Ta and water vapor pressure. In general, the generated L_d dataset has a higher spatial resolution and accuracy, which can contribute to perfect the existing radiation products.

1. Introduction

The surface downward longwave radiation (L_d, 4–100 μm) is an indispensable component needed to study the Earth’s surface radiation budget and energy balance [1]. Currently, there are four main ways of obtaining L_d: ground measurement data, reanalysis retrieval methods, general circulation model (GCM) simulations and satellite products. However, L_d is not always treated as a conventional observation as other common meteorological parameters are, such as air temperature (Ta), relative humidity (RH), etc. Moreover, its observation stations are sparsely distributed and even entirely absent in certain areas due to a high cost, a difficult calibration process, and a required quality control step [2,3,4,5]. In addition, there are uncertainties and biases in GCM simulations [6,7,8], reanalysis retrievals [9,10], and satellite products [11]. Therefore, establishing a more accurate long-term global L_d dataset is not only useful for improving the knowledge of the surface radiation balance but is also helpful for perfecting the existing L_d products.

Under clear-sky conditions, L_d is primarily influenced by temperature profiles and water vapor in the lower atmosphere. Zeppetello et al. [12] found that L_d is tightly coupled to surface temperature, and changes in surface temperature cause at least 63% of the clear-sky L_d response in greenhouse forcing. Water vapor is the most crucial atmospheric gas contributing to thermal radiation which can absorb and emit longwave radiation, thereby resulting in L_d estimates with great uncertainty [13]. RH, which is closely related to water vapor pressure, is the percentage of water vapor pressure in the atmosphere to the saturated vapor pressure at a given temperature. Numerous studies [5,14,15,16,17,18,19,20] have estimated L_d on the basis of traditional methods using Ta, water vapor, RH, and other basic variables derived from meteorological observations. These methods mainly include empirical, physics-based, and hybrid methods. Among them, empirical models, including the representative Brunt [14] and Brutsaert [15] equations, establish the regression relationship between various meteorological parameters and L_d observations, with an accuracy that is mainly limited by ground measurements and actual geographical environments, such as climate and terrain. Although this method is relatively simple, it is difficult to apply to L_d estimation on a large regional scope. Compared with empirical methods, physics-based methods containing the LOWTRAN and MODTRAN models can not only estimate L_d with a high accuracy but also describe the atmospheric radiative transfer process in detail [21,22,23]. Due to the intricacy of the model and the difficulty associated with obtaining an input dataset, however, this approach is only used for research and is difficult to apply to business products [4]. Hybrid methods [24,25,26,27,28] establish the relationship between L_d and the top-of-atmosphere radiance on the basis of physical radiative transfer processes. In contrast, this method, with its higher simulation accuracy and greater general applicability, can be applied on a global scale, which has become an effective method for L_d retrieval. For example, Wang et al. [27] developed a hybrid method to estimate instantaneous land clear L_d on the basis of extensive radiative transfer simulation and statistical analysis, obtaining root mean squared error (RMSE) values of 17.60 W m⁻² (Terra) and 16.17 W m⁻² (Aqua) for the nonlinear models.

Under cloudy conditions, the influence of clouds on L_d is also nonnegligible. Clouds are visible polymers of tiny water droplets or ice crystals formed by the condensation of water vapor in the atmosphere, which can absorb heat from the ground and radiate it back to the surface to enhance L_d [29,30]. The cloud cover fraction is mostly utilized to quantify the effects of clouds on L_d and is an essential parameter for L_d estimation under cloudy conditions, which can be obtained from ground measurements and satellite cloud detection products [31,32,33]. However, the effects of clouds cannot be corrected when cloud cover fraction observations are not available. Crawford et al. [34] first proposed that the cloud cover fraction under cloudy-sky conditions can be estimated from the proportion of the observed surface downward shortwave radiation (S_d) to the theoretical clear-sky S_d under the same conditions. They evaluated the performance of estimating L_d using S_d, barometric pressure, vapor pressure, and temperature datasets. The evaluation results showed that the RMSEs and mean bias errors (MBE) of the monthly L_d estimates ranged from 11 to 22 W m⁻² and −9 to 4 W m⁻² compared to ground observations over a one-year time period, respectively, which indicated that it is reliable to use S_d to represent the impact of clouds on L_d. It is easier to obtain S_d data compared with cloud cover fractions, so an increasing number of studies have utilized S_d to estimate L_d under cloudy conditions [5,13,35,36,37]. Choi et al. [35] estimated the daily L_d using 2-m air temperature, 2-m RH, and S_d observations in Florida from 2004 to 2005, obtaining RMSEs of less than 13 W m⁻² and squared correlation coefficients (R²) of more than 0.9 relative to the ground measurements collected at 11 stations. Lhomme et al. [13] demonstrated that the cloud correction function of the Crawford et al. [34] model also performed relatively credibly for estimating L_d in high elevation regions between 3700 and 4100 m above sea level. The presence of clouds makes it impossible for satellites to accurately observe surface information. It is also difficult to model the properties of clouds due to the uncertainty associated with their distribution and variability. The ready availability of S_d data makes the L_d estimation model more readily applicable under cloudy conditions.

In addition, S_d and L_d both show a strong dependence on altitude. Zeng et al. [38] evaluated the global land surface satellite (GLASS) L_d product using the ground observations collected from 141 stations in six networks at different surface elevations. The RMSE values are 22.09, 23.31, 26.94, and 26.99 W m⁻² at elevations of <500, 500–1000, 1000–3000, and >3000 m, respectively. The bias values are −3.19, −4.73, −2.26, and 15.34 W m⁻² at the four elevation intervals, respectively. The validation results showed that the performance of L_d degraded as the surface elevation increased. This may be due to special environmental conditions present at high altitudes with lower air pressures, smaller water vapor densities, and fewer clouds, leading to a greater uncertainty in S_d and L_d data at high elevations [37,39,40,41,42,43]. In addition, some studies have also quantitatively measured the effect of elevation on L_d and attempted to correct its deviation [37,39,40,41,42]. Yang et al. [42] reported that the MBE of GEWEX-SRB V2.5 L_d can be reduced by 7–10 W m⁻² after an altitudinal correction of 2.8 W m⁻² per hundred meters in the Tibet Plateau. It can be concluded that the influence of elevation cannot be ignored in addition to the abovementioned influencing factors including temperature profiles, water vapor, and clouds. Although the importance of elevation has been verified by previous studies, few studies have taken elevation as an important variable to predict L_d. This paper used elevation as the input variable of the model, hoping to reduce the errors caused by elevation.

Based on the above summary, it is clear that L_d is closely related to Ta, RH, water vapor, S_d, and elevation. Therefore, this study utilized the gradient boosting regression tree (GBRT) method with the daily mean Ta of 2 m, RH at 1000 hPa, total column water vapor, S_d, and elevation to estimate daily L_d over global land surface from 2000 to 2018. In contrast to prior methods, this machine learning method can automatically establish the relationship between the input data and target variable, and has a strong predictive ability [44,45], which has been widely employed to retrieve radiation [46,47,48,49]. Yang et al. [46] applied the GBRT method to estimate daily S_d with a spatial resolution of 5 km in China using ground observations and satellite retrievals with good results. The RMSE and R between the ground measurements and daily L_d estimates were 27.71 W m⁻² and 0.91, respectively, under cloudy conditions; these values were 42.97 W m⁻² and 0.80, respectively, under clear conditions. To date, few studies have used this method to predict L_d over the globe based on ground observations. We demonstrated that it can be reasonably and reliably used for L_d estimation by building the relationship between L_d observations and its influencing factors based on the GBRT method [49,50]. Therefore, the objective of this study is to use the GBRT model to generate a 5-km L_d dataset over the global land surface with a daily time scale from 2000 to 2018.

The structure of this paper is as follows: Section 2 introduces the data used, including the ground measurements, ERA5 reanalysis data, GLASS S_d, global multi-resolution terrain elevation data 2010, and existing L_d products. The detailed model construction process is displayed and described in Section 3. Section 4 provides the evaluation results and analyzes the spatiotemporal distribution of L_d. Finally, the discussion and conclusion are presented in Section 5 and Section 6, respectively.

2. Data

2.1. Ground Measurements

The ground measurements of surface downward longwave radiation (L_d) used in this study from 2000 to 2018 were collected from the AmeriFlux network (175 sites), AsiaFlux network (26 sites), baseline surface radiation network (BSRN, 57 sites), surface radiation budget network (SURFRAD, 7 sites), and FLUXNET (84 sites). The observation sites were randomly divided into 90% (314 sites) and 10% (35 sites) datasets, as shown in Figure 1. After removing the outliers, the L_d observations collected at 314 sites were used as target variables to build and train the model. The remaining L_d observations collected at 35 sites were used to evaluate the generated global land daily L_d. The spatial distribution of the observation sites used to build the model and validate it is shown in Figure 1. The detailed information of ground sites is listed in the Appendix A 

Table A1. Detailed information of the ground sites.

Number	Site Code	Site Name	Latitude (deg)	Longitude (deg)	Elevation (m)	Time Period
1	BR-Npw	Northern Pantanal Wetland	−16.50	−56.41	120	2013–2017
2	BR-Sa3	Santarem-Km83-Logged Forest	−3.02	−54.97	100	2001–2004
3	CA-ARF	Attawapiskat River Fen	52.70	−83.96	88	2011–2015
4	CA-Ca1	British Columbia–1949 Douglas-fir stand	49.87	−125.33	300	2000–2010
5	CA-Ca3	British Columbia–Pole sapling Douglas-fir stand	49.53	−124.90	\	2002–2016
6	CA-Cbo	Ontario–Mixed Deciduous, Borden Forest Site	44.32	−79.93	120	2005–2018
7	CA-DBB	Delta Burns Bog	49.13	−122.98	4	2016–2018
8	CA-Gro	Ontario–Groundhog River, Boreal Mixedwood Forest	48.22	−82.16	340	2003–2014
9	CA-Na1	New Brunswick–1967 Balsam Fir–Nashwaak Lake Site 01 (Mature balsam fir forest)	46.47	−67.10	341	2003–2005
10	CA-Oas	Saskatchewan–Western Boreal, Mature Aspen	53.63	−106.20	530	2000–2010
11	CA-Obs	Saskatchewan–Western Boreal, Mature Black Spruce	53.99	−105.12	628.94	2000–2010
12	CA-Ojp	Saskatchewan–Western Boreal, Mature Jack Pine	53.92	−104.69	579	2000–2010
13	CA-Qcu	Quebec–Eastern Boreal, Black Spruce/Jack Pine Cutover	49.27	−74.04	392.3	2004–2010
14	CA-Qfo	Quebec–Eastern Boreal, Mature Black Spruce	49.69	−74.34	382	2003–2010
15	CA-SCB	Scotty Creek Bog	61.31	−121.30	280	2014–2017
16	CA-SCC	Scotty Creek Landscape	61.31	−121.30	285	2013–2016
17	CA-SF1	Saskatchewan–Western Boreal, forest burned in 1977	54.49	−105.82	536	2003–2006
18	CA-SF2	Saskatchewan–Western Boreal, forest burned in 1989	54.25	−105.88	520	2002–2005
19	CA-SF3	Saskatchewan–Western Boreal, forest burned in 1998	54.09	−106.01	540	2002–2006
20	CA-SJ1	Saskatchewan–Western Boreal, Jack Pine forest harvested in 1994	53.91	−104.66	580	2001–2010
21	CA-SJ2	Saskatchewan–Western Boreal, Jack Pine forest harvested in 2002	53.95	−104.65	580	2003–2010
22	CA-SJ3	Saskatchewan–Western Boreal, Jack Pine forest harvested in 1975 (BOREAS Young Jack Pine)	53.88	−104.65	\	2004–2010
23	CA-TP4	Ontario–Turkey Point 1939 Plantation White Pine	42.71	−80.36	184	2003–2017
24	CA-TPD	Ontario–Turkey Point Mature Deciduous	42.64	−80.56	260	2012–2017
25	CA-WP1	Alberta–Western Peatland–LaBiche River,Black Spruce/Larch Fen	54.95	−112.47	540	2003–2009
26	US-A03	ARM-AMF3-Oliktok	70.50	−149.88	5	2014–2018
27	US-A10	ARM-NSA-Barrow	71.32	−156.61	4	2011–2018
28	US-A32	ARM-SGP Medford hay pasture	36.82	−97.82	335	2015–2017
29	US-A74	ARM SGP milo field	36.81	−97.55	337	2016–2017
30	US-AR1	ARM USDA UNL OSU Woodward Switchgrass 1	36.43	−99.42	611	2009–2012
31	US-AR2	ARM USDA UNL OSU Woodward Switchgrass 2	36.64	−99.60	646	2009–2012
32	US-ARM	ARM Southern Great Plains site- Lamont	36.61	−97.49	314	2003–2018
33	US-An1	Anaktuvuk River Severe Burn	68.99	−150.28	600	2008–2009
34	US-An2	Anaktuvuk River Moderate Burn	68.95	−150.21	600	2008–2019
35	US-An3	Anaktuvuk River Unburned	68.93	−150.27	600	2008–2010
36	US-Bi1	Bouldin Island Alfalfa	38.10	−121.50	−2.7	2016–2018
37	US-Bi2	Bouldin Island corn	38.11	−121.54	−5	2017–2018
38	US-Bkg	Brookings	44.35	−96.84	510	2004–2010
39	US-Blk	Black Hills	44.16	−103.65	1718	2004–2008
40	US-Bo1	Bondville	40.01	−88.29	219	2000–2008
41	US-Br1	Brooks Field Site 10- Ames	41.97	−93.69	313	2005–2011
42	US-Br3	Brooks Field Site 11- Ames	41.97	−93.69	313	2005–2011
43	US-CPk	Chimney Park	41.07	−106.12	2750	2009–2013
44	US-ChR	Chestnut Ridge	35.93	−84.33	286	2005–2010
45	US-Ctn	Cottonwood	43.95	−101.85	744	2006–2009
46	US-Dia	Diablo	37.68	−121.53	323	2010–2012
47	US-Dk1	Duke Forest-open field	35.97	−79.09	168	2004–2008
48	US-Dk2	Duke Forest-hardwoods	35.97	−79.10	168	2004–2008
49	US-Dk3	Duke Forest–loblolly pine	35.98	−79.09	163	2004–2008
50	US-EDN	Eden Landing Ecological Reserve	37.62	−122.11	\	2018
51	US-EML	Eight Mile Lake Permafrost thaw gradient, Healy Alaska.	63.88	−149.25	700	2011–2018
52	US-FPe	Fort Peck	48.31	−105.10	634	2000–2008
53	US-FR2	Freeman Ranch- Mesquite Juniper	29.95	−98.00	271.9	2008
54	US-FR3	Freeman Ranch- Woodland	29.94	−97.99	232	2008–2012
55	US-Fmf	Flagstaff–Managed Forest	35.14	−111.73	2160	2005–2010
56	US-Fuf	Flagstaff–Unmanaged Forest	35.09	−111.76	2180	2005–2010
57	US-Fwf	Flagstaff–Wildfire	35.45	−111.77	2270	2005–2010
58	US-GLE	GLEES	41.37	−106.24	3197	2004–2018
59	US-Goo	Goodwin Creek	34.25	−89.87	87	2002–2006
60	US-HBK	Hubbard Brook Experimental Forest	43.94	−71.72	367	2017–2018
61	US-HRA	Humnoke Farm Rice Field–Field A	34.59	−91.75	\	2016–2017
62	US-HRC	Humnoke Farm Rice Field–Field C	34.59	−91.75	\	2016–2017
63	US-Ha2	Harvard Forest Hemlock Site	42.54	−72.18	360	2014–2018
64	US-Hn3	Hobcaw Barony Longleaf Pine Restoration	46.69	−119.46	120.9	2017–2018
65	US-Ho1	Howland Forest (main tower)	45.20	−68.74	60	2007–2018
66	US-Ho2	Howland Forest (west tower)	45.21	−68.75	61	2007–2009
67	US-Ho3	Howland Forest (harvest site)	45.21	−68.73	61	2007–2009
68	US-Ivo	Ivotuk	68.49	−155.75	568	2003–2006
69	US-KFS	Kansas Field Station	39.06	−95.19	310	2008–2018
70	US-KLS	Kansas Land Institute	38.77	−97.57	373	2012–2017
71	US-KM4	KBS Marshall Farms Smooth Brome Grass (Ref)	42.44	−85.33	288	2010–2018
72	US-KS3	Kennedy Space Center (salt marsh)	28.71	−80.74	0	2018
73	US-KUT	KUOM Turfgrass Field	44.99	−93.19	301	2006–2009
74	US-Kon	Konza Prairie LTER (KNZ)	39.08	−96.56	417	2006–2018
75	US-Los	Lost Creek	46.08	−89.98	480	2014–2018
76	US-MMS	Morgan Monroe State Forest	39.32	−86.41	275	2000–2018
77	US-MOz	Missouri Ozark Site	38.74	−92.20	219.4	2004–2017
78	US-MRf	Mary’s River (Fir) site	44.65	−123.55	263	2007–2011
79	US-MSR	Montana Sun River winter wheat	47.48	−111.72	1110	2016
80	US-Me2	Metolius mature ponderosa pine	44.45	−121.56	1253	2005–2018
81	US-Me3	Metolius-second young aged pine	44.32	−121.61	1005	2009
82	US-Me6	Metolius Young Pine Burn	44.32	−121.61	998	2010–2018
83	US-Men	Lake Mendota, Center for Limnology Site	43.08	−89.40	260	2012–2018
84	US-Mpj	Mountainair Pinyon-Juniper Woodland	34.44	−106.24	2196	2008–2018
85	US-MtB	Mt Bigelow	32.42	−110.73	2573	2009–2018
86	US-NC1	Mt Bigelow	35.81	−76.71	5	2005–2012
87	US-NC2	NC_Loblolly Plantation	35.80	−76.67	5	2005–2018
88	US-NC3	NC_Clearcut#3	35.80	−76.66	5	2013–2018
89	US-NC4	NC_AlligatorRiver	35.79	−75.90	1	2015–2018
90	US-NGB	NGEE Arctic Barrow	71.28	−156.61	5.273	2012–2018
91	US-NGC	NGEE Arctic Council	64.86	−163.70	35	2017–2018
92	US-NR1	Niwot Ridge Forest (LTER NWT1)	40.03	−105.55	3050	2000–2018
93	US-Ne1	Mead–irrigated continuous maize site	41.17	−96.48	361	2001–2018
94	US-Ne2	Mead–irrigated maize-soybean rotation site	41.16	−96.47	362	2001–2018
95	US-Ne3	Mead–rainfed maize-soybean rotation site	41.18	−96.44	363	2001–2018
96	US-Orv	Olentangy River Wetland Research Park	40.02	−83.02	221	2011–2016
97	US-Oho	Oak Openings	41.55	−83.84	230	2004–2013
98	US-PHM	Plum Island High Marsh	42.74	−70.83	1.4	2013–2018
99	US-Pnp	Lake Mendota, Picnic Point Site	43.09	−89.42	260	2016–2018
100	US-Prr	Poker Flat Research Range Black Spruce Forest	65.12	−147.49	210	2010–2016
101	US-Rls	RCEW Low Sagebrush	43.14	−116.74	1608	2014–2018
102	US-Rms	RCEW Mountain Big Sagebrush	43.06	−116.75	2111	2014–2018
103	US-Ro1	Rosemount- G21	44.71	−93.09	260	2004–2016
104	US-Ro2	Rosemount- C7	44.73	−93.09	292	2015–2016
105	US-Ro4	Rosemount Prairie	44.68	−93.07	274	2015–2018
106	US-Ro5	Rosemount I18_South	44.69	−93.06	283	2017–2018
107	US-Ro6	Rosemount I18_North	44.69	−93.06	282	2017–2018
108	US-Rpf	Poker Flat Research Range: Succession from fire scar to deciduous forest	65.12	−147.43	497	2013–2018
109	US-Rwe	RCEW Reynolds Mountain East	43.07	−116.76	2098	2005–2007
110	US-Rwf	RCEW Upper Sheep Prescibed Fire	43.12	−116.72	1878	2014–2018
111	US-Rws	Reynolds Creek Wyoming big sagebrush	43.17	−116.71	1425	2014–2018
112	US-SFP	Sioux Falls Portable	43.24	−96.90	386	2007–2009
113	US-SRC	Santa Rita Creosote	31.91	−110.84	950	2008–2014
114	US-SRG	Santa Rita Grassland	31.79	−110.83	1291	2008–2018
115	US-SRM	Santa Rita Mesquite	31.82	−110.87	1120	2004–2018
116	US-Seg	Sevilleta grassland	34.36	−106.70	1622	2007–2018
117	US-Ses	Sevilleta shrubland	34.33	−106.74	1604	2007–2018
118	US-Skr	Shark River Slough (Tower SRS-6) Everglades	25.36	−81.08	0	2004–2011
119	US-Slt	Silas Little- New Jersey	39.91	−74.60	30	2007–2012
120	US-Sne	Sherman Island Restored Wetland	38.04	−121.75	−5	2016–2018
121	US-Snf	Sherman Barn	38.04	−121.73	−4	2018
122	US-Srr	Suisun marsh–Rush Ranch	38.20	−122.03	8	2014–2017
123	US-Ton	Tonzi Ranch	38.43	−120.97	177	2014–2018
124	US-Tw1	Twitchell Wetland West Pond	38.11	−121.65	−5	2011–2018
125	US-Tw2	Twitchell Corn	38.10	−121.64	−5	2012–2013
126	US-Tw3	Twitchell Alfalfa	38.12	−121.65	−4	2013–2018
127	US-Tw4	Twitchell East End Wetland	38.10	−121.64	−5	2013–2018
128	US-Tw5	East Pond Wetland	38.11	−121.64	−5	2018
129	US-UM3	Douglas Lake	45.57	−84.67	234	2013–2014
130	US-UMB	Univ. of Mich. Biological Station	45.56	−84.71	234	2007–2018
131	US-UMd	UMBS Disturbance	45.56	−84.70	239	2008–2018
132	US-Uaf	University of Alaska, Fairbanks	64.87	−147.86	155	2009–2018
133	US-UiA	University of Illinois Switchgrass	40.06	−88.20	224	2015
134	US-Var	Vaira Ranch- Ione	38.41	−120.95	129	2004–2018
135	US-Vcm	Valles Caldera Mixed Conifer	35.89	−106.53	3003	2009–2018
136	US-Vcp	Valles Caldera Ponderosa Pine	35.86	−106.60	2542	2007–2018
137	US-Vcs	Valles Caldera Sulphur Springs Mixed Conifer	35.92	−106.61	2752	2016–2018
138	US-WBW	Walker Branch Watershed	35.96	−84.29	283	2001–2007
139	US-WCr	Willow Creek	45.81	−90.08	520	2000–2018
140	US-WPT	Winous Point North Marsh	41.46	−83.00	175	2011–2013
141	US-Wdn	Walden	40.78	−106.26	2469	2006–2008
142	US-Wgr	Willamette Grass	45.11	−122.66	52	2015
143	US-Whs	Walnut Gulch Lucky Hills Shrub	31.74	−110.05	1370	2009–2018
144	US-Wjs	Willard Juniper Savannah	34.43	−105.86	1931	2007–2018
145	US-Wkg	Walnut Gulch Kendall Grasslands	31.74	−109.94	1531	2004–2018
146	US-Wpp	Willamette Poplar	44.14	−123.18	111	2015
147	US-Wrc	Wind River Crane Site	45.82	−121.95	371	2000–2015
148	US-xAB	NEON Abby Road (ABBY)	45.76	−122.33	363	2017–2018
149	US-xBN	NEON Caribou Creek–Poker Flats Watershed (BONA)	65.15	−147.50	263	2018
150	US-xBR	NEON Bartlett Experimental Forest (BART)	44.06	−71.29	232	2017–2018
151	US-xCP	NEON Central Plains Experimental Range (CPER)	40.82	−104.75	1654	2017–2018
152	US-xDC	NEON Dakota Coteau Field School (DCFS)	47.16	−99.11	559	2017–2018
153	US-xDJ	NEON Delta Junction (DEJU)	63.88	−145.75	529	2017–2018
154	US-xDL	NEON Dead Lake (DELA)	32.54	−87.80	22	2017–2018
155	US-xGR	NEON Great Smoky Mountains National Park, Twin Creeks (GRSM)	35.69	−83.50	579	2018
156	US-xHA	NEON Harvard Forest (HARV)	42.54	−72.17	351	2017–2018
157	US-xHE	NEON Healy (HEAL)	63.88	−149.21	705	2017–2018
158	US-xJE	NEON Jones Ecological Research Center (JERC)	31.19	−84.47	44	2017–2018
159	US-xJR	NEON Jornada LTER (JORN)	32.59	−106.84	1329	2017–2018
160	US-xKA	NEON Konza Prairie Biological Station–Relocatable (KONA)	39.11	−96.61	1329	2017–2018
161	US-xKZ	NEON Konza Prairie Biological Station (KONZ)	39.10	−96.56	381	2017–2018
162	US-xNG	NEON Northern Great Plains Research Laboratory (NOGP)	46.77	−100.92	578	2017–2018
163	US-xNQ	NEON Onaqui-Ault (ONAQ)	40.18	−112.45	1685	2017–2018
164	US-xRM	NEON Rocky Mountain National Park, CASTNET (RMNP)	40.28	−105.55	2743	2017–2018
165	US-xSE	NEON Smithsonian Environmental Research Center (SERC)	38.89	−76.56	15	2017–2018
166	US-xSL	NEON North Sterling, CO (STER)	40.46	−103.03	1364	2017–2018
167	US-xSP	NEON Soaproot Saddle (SOAP)	37.03	−119.26	1160	2017–2018
168	US-xSR	NEON Santa Rita Experimental Range (SRER)	31.91	−110.84	983	2017–2018
169	US-xST	NEON Steigerwaldt Land Services (STEI)	45.51	−89.59	481	2017–2018
170	US-xTE	NEON Lower Teakettle (TEAK)	37.01	−119.01	2147	2018
171	US-xTR	NEON Treehaven (TREE)	45.49	−89.59	472	2017–2018
172	US-xUK	NEON The University of Kansas Field Station (UKFS)	39.04	−95.19	335	2017–2018
173	US-xUN	NEON University of Notre Dame Environmental Research Center (UNDE)	46.23	−89.54	518	2017–2018
174	US-xWD	NEON Woodworth (WOOD)	47.13	−99.24	579	2017–2018
175	US-xWR	NEON Wind River Experimental Forest (WREF)	45.82	−121.95	407	2018
176	MSE	Mase paddy flux site	36.05	140.03	13	2001
177	PSO	Pasoh Forest Reserve	2.97	102.31	75–150	2003–2009
178	BKS	Bukit Soeharto	-0.86	117.04	20	2001–2002
179	CBS	Changbaishan Site	41.40	128.10	731	2003–2005
180	FHK	Fuji Hokuroku Flux Observation Site	35.44	138.76	1050–1150	2006–2012
181	GCK	Gwangreung Coniferous forest	37.75	127.16	132	2007–2008
182	HBG	Haibei Potentilla fruticisa bosk Site	37.48	101.20	756	2003–2004
183	HFK	Haenam Farmland	34.55	127.57	12	2008
184	IRI	IRRI Flux Research Site	14.14	121.27	21	2009–2014
185	KBU	Kherlenbayan Ulaan	47.21	108.74	1235	2003–2009
186	LSH	Laoshan	45.28	127.58	340	2002
187	MBF	Moshiri Birch Forest Site	44.38	142.32	585	2003–2005
188	MKL	Mae Klong	14.59	98.84	585	2003–2004
189	MMF	Moshiri Mixd Forest Site	44.32	142.26	340	2003–2005
190	PDF	Palangkaraya drained forest	−2.35	114.04	30	2002–2005
191	QYZ	Qianyanzhou Site	26.73	115.07	100	2003–2004
192	SKR	Sakaerat	14.49	101.92	543	2001–2003
193	SKT	Southern Khentei Taiga	48.35	108.65	1630	2003–2006
194	SMF	Seto Mixed Forest Site	35.26	137.08	205	2002–2015
195	SWL	Suwa Lake Site	36.05	138.11	759	2015–2018
196	TKC	Takayama evergreen coniferous forest site	36.14	137.37	800	2007
197	TMK	Tomakomai Flux Research Site	42.74	141.51	140	2001–2003
198	TSE	CC-LaG Teshio Experimental Forest	45.06	142.11	70	2001–2005
199	YCS	Yuchen Site	36.83	116.57	28	2003–2005
200	YLF	Yakutsk Spasskaya Pad larch	62.26	129.17	220	2003–2007
201	YPF	Yakutsk Pine	62.24	129.65	220	2004–2007
202	ALE	Alert	82.49	−62.42	127	2004–2014
203	ASP	Alice Springs	−23.80	133.89	547	2000–2018
204	BAR	Barrow	71.32	−156.61	8	2000–2017
205	BIL	Billings	36.61	−97.52	317	2000–2017
206	BON	Bondville	40.07	−88.37	213	2009–2018
207	BOS	Boulder	40.13	−105.24	1689	2009–2018
208	BOU	Boulder	40.05	−105.01	1577	2000–2016
209	BRB	Brasilia	−15.60	−47.71	1023	2008–2018
210	CAB	Cabauw	51.97	4.93	0	2005–2018
211	CAM	Camborne	50.22	−5.32	88	2001–2017
212	CAR	Carpentras	44.08	5.06	100	2000–2018
213	CNR	Cener	42.82	−1.60	471	2009–2018
214	COC	Cocos Island	−12.19	96.84	6	2004–2018
215	DAA	De Aar	−30.67	23.99	1287	2000–2018
216	DAR	Darwin	−12.43	130.89	30	2002–2015
217	DOM	Concordia Station, Dome C	−75.10	123.38	3233	2006–2018
218	DRA	Desert Rock	36.63	−116.02	1007	2009–2018
219	DWN	Darwin Met Office	−12.42	130.89	32	2008–2018
220	E13	Southern Great Plains	36.61	−97.49	318	2000–2017
221	ENA	Eastern North Atlantic	39.09	−28.03	15.2	2013–2015
222	EUR	Eureka	79.99	−85.94	85	2007–2011
223	FLO	Florianopolis	−27.60	−48.52	11	2013–2018
224	FPE	Fort Peck	48.32	−105.10	634	2009–2018
225	FUA	Fukuoka	33.58	130.38	3	2010–2018
226	GAN	Gandhinagar	23.11	72.63	65	2014–2015
227	GCR	Goodwin Creek	34.25	−89.87	98	2009–2018
228	GOB	Gobabeb	−23.56	15.04	407	2012–2018
229	GUR	Gurgaon	28.42	77.16	259	2014–2018
230	GVN	Georg von Neumayer	−70.65	−8.25	42	2000–2018
231	HOW	Howrah	22.55	88.31	51	2014–2018
232	ISH	Ishigakijima	24.34	124.16	5.7	2010–2018
233	LAU	Lauder	−45.05	169.69	350	2000–2018
234	LER	Lerwick	60.14	−1.18	80	2001–2017
235	LIN	Lindenberg	52.21	14.12	125	2000–2017
236	LRC	Langley Research Center	37.10	−76.39	3	2014–2018
237	LYU	Lanyu Station	22.04	121.56	324	2018
238	MAN	Momote	−2.06	147.43	6	2000–2013
239	NAU	Nauru Island	−0.52	166.92	7	2000–2013
240	NEW	Newcastle	−32.88	151.73	18.5	2017–2018
241	NYA	Ny-Ålesund	78.93	11.93	11	2000–2018
242	PAL	Palaiseau, SIRTA Observatory	48.71	2.21	156	2003–2018
243	PAY	Payerne	46.82	6.94	491	2000–2018
244	PSU	Rock Springs	40.72	−77.93	376	2009–2018
245	PTR	Petrolina	−9.07	−40.32	387	2008–2018
246	REG	Regina	50.21	−104.71	578	2000–2011
247	SAP	Sapporo	43.06	141.33	17.2	2010–2018
248	SBO	Sede Boqer	30.86	34.78	500	2003–2012
249	SMS	São Martinho da Serra	−29.44	−53.82	489	2008–2017
250	SON	Sonnblick	47.05	12.96	3108.9	2013–2018
251	SOV	Solar Village	24.91	46.41	650	2000–2002
252	SXF	Sioux Falls	43.73	−96.62	473	2009–2018
253	SYO	Syowa	−69.01	39.59	18	2000–2018
254	TAM	Tamanrasset	22.79	5.53	1385	2000–2018
255	TAT	Tateno	36.06	140.13	25	2000–2018
256	TIR	Tiruvallur	13.09	79.97	36	2014–2018
257	TOR	Toravere	58.25	26.46	70	2003–2018
258	XIA	Xianghe	39.75	116.96	32	2005–2015
259	AT-Neu	Neustift	47.12	11.32	970	2005–2012
260	AU-ASM	Alice Springs	−22.28	133.25	\	2010–2014
261	AU-Ade	Adelaide River	−13.08	131.12	\	2007–2009
262	AU-Cpr	Calperum	−34.00	140.59	\	2010–2014
263	AU-Cum	Cumberland Plain	−33.62	150.72	\	2012–2014
264	AU-DaP	Daly River Savanna	−14.06	131.32	\	2007–2013
265	AU-DaS	Daly River Cleared	−14.16	131.39	\	2008–2014
266	AU-Dry	Dry River	−15.26	132.37	\	2008–2014
267	AU-Emr	Emerald	−23.86	148.47	\	2011–2013
268	AU-Fog	Fogg Dam	−12.55	131.31	\	2006–2008
269	AU-GWW	Great Western Woodlands, Western Australia, Australia	−30.19	120.65	\	2013–2014
270	AU-Gin	Gingin	−31.38	115.71	\	2011–2014
271	AU-Lox	Loxton	−34.47	140.66	\	2008–2009
272	AU-RDF	Red Dirt Melon Farm, Northern Territory	−14.56	132.48	\	2011–2013
273	AU-Rig	Riggs Creek	−36.65	145.58	\	2011–2014
274	AU-Rob	Robson Creek, Queensland, Australia	−17.12	145.63	\	2014
275	AU-Stp	Sturt Plains	−17.15	133.35	\	2008–2014
276	AU-TTE	Ti Tree East	−22.29	133.64	\	2012–2014
277	AU-Tum	Tumbarumba	−35.66	148.15	1200	2007–2014
278	AU-Whr	Whroo	−36.67	145.03	\	2011–2014
279	AU-Wom	Wombat	−37.42	144.09	705	2010–2014
280	AU-Ync	Jaxa	−34.99	146.29	\	2012–2014
281	BE-Bra	Brasschaat	51.31	4.52	16	2007–2014
282	BE-Lon	Lonzee	50.55	4.75	167	2005–2014
283	CH-Cha	Chamau	47.21	8.41	393	2005–2014
284	CH-Dav	Davos	46.82	9.86	1639	2006–2014
285	CH-Fru	Früebüel	47.12	8.54	982	2005–2014
286	CH-Lae	Laegern	47.48	8.36	689	2005–2014
287	CH-Oe1	Oensingen grassland	47.29	7.73	450	2003–2008
288	CH-Oe2	Oensingen crop	47.29	7.73	452	2004–2014
289	CN-Cha	Changbaishan	42.40	128.10	\	2003–2005
290	CN-Cng	Changling	44.59	123.51	\	2007–2010
291	CN-Dan	Dangxiong	30.50	91.07	\	2004–2005
292	CN-Din	Dinghushan	23.17	112.54	\	2003–2005
293	CN-Ha2	Haibei Shrubland	37.61	101.33	\	2003–2005
294	CN-Qia	Qianyanzhou	26.74	115.06	\	2003–2005
295	CZ-wet	Trebon (CZECHWET)	49.02	14.77	426	2006–2014
296	DE-Akm	Anklam	53.87	13.68	−1	2009–2014
297	DE-Geb	Gebesee	51.10	10.91	161.5	2001–2014
298	DE-Gri	Grillenburg	50.95	13.51	385	2006–2014
299	DE-Hai	Hainich	51.08	10.45	430	2002–2012
300	DE-Kli	Klingenberg	50.89	13.52	478	2004–2014
301	DE-Lkb	Lackenberg	49.10	13.30	1308	2009–2013
302	DE-Lnf	Leinefelde	51.33	10.37	451	2002–2012
303	DE-Obe	Oberbärenburg	50.79	13.72	734	2008–2014
304	DE-RuR	Rollesbroich	50.62	6.30	514.7	2011–2014
305	DE-RuS	Selhausen Juelich	50.87	6.45	102.755	2011–2014
306	DE-SfN	Schechenfilz Nord	47.81	11.33	590	2012–2014
307	DE-Spw	Spreewald	51.89	14.03	61	2010–2014
308	DE-Tha	Tharandt	50.96	13.57	385	2004–2014
309	DE-Zrk	Zarnekow	53.88	12.89	0	2013–2014
310	FI-Hyy	Hyytiala	61.85	24.29	181	2009–2014
311	FI-Lom	Lompolojankka	68.00	24.21	274	2007–2009
312	FR-Gri	Grignon	48.84	1.95	125	2004–2014
313	FR-LBr	Le Bray	44.72	−0.77	61	2003–2008
314	FR-Pue	Puechabon	43.74	3.60	270	2005–2014
315	IT-BCi	Borgo Cioffi	40.52	14.96	20	2006–2011
316	IT-CA1	Castel d’Asso1	42.38	12.03	200	2011–2014
317	IT-CA2	Castel d’Asso2	42.38	12.03	200	2011–2014
318	IT-CA3	Castel d’Asso3	42.38	12.02	197	2011–2014
319	IT-Col	Collelongo	41.85	13.59	1560	2004–2014
320	IT-Isp	Ispra ABC-IS	45.81	8.63	210	2013–2014
321	IT-La2	Lavarone2	45.95	11.29	1350	2000–2002
322	IT-Lav	Lavarone	45.96	11.28	1353	2003–2004
323	IT-MBo	Monte Bondone	46.01	11.05	1550	2003–2013
324	IT-Noe	Arca di Noe–Le Prigionette	40.61	8.15	25	2004–2014
325	IT-Ren	Renon	46.59	11.43	1730	2003–2013
326	IT-Ro2	Roccarespampani 2	42.39	11.92	160	2010–2012
327	IT-SR2	San Rossore 2	43.73	10.29	4	2013–2014
328	IT-SRo	San Rossore	43.73	10.28	6	2004–2008
329	IT-Tor	Torgnon	45.84	7.58	2160	2008–2014
330	JP-MBF	Moshiri Birch Forest Site	44.39	142.32	\	2003–2005
331	NL-Hor	Horstermeer	52.24	5.07	2.2	2004–2011
332	NL-Loo	Loobos	52.17	5.74	25	2000–2014
333	RU-Che	Cherski	68.61	161.34	6	2002–2005
334	RU-Fyo	Fyodorovskoye	56.46	32.92	265	2000–2014
335	SE-St1	Stordalen grassland	68.35	19.05	351	2012–2014
336	SJ-Blv	Bayelva, Spitsbergen	78.92	11.83	25	2008–2009
337	US-CRT	Curtice Walter-Berger cropland	41.63	−83.35	180	2011–2013
338	US-GBT	GLEES Brooklyn Tower	41.37	−106.24	3191	2000–2006
339	US-Syv	Sylvania Wilderness Area	46.24	−89.35	540	2012–2014
340	US-Tw4	Twitchell East End Wetland	38.10	−121.64	−5	2013–2014
341	ZA-Kru	Skukuza	−25.02	31.50	359	2000–2003
342	ZM-Mon	Mongu	−15.44	23.25	1053	2000–2009
343	BND	Bondville	40.05	−88.37	230	2000–2018
344	DRA	Desert Rock	36.62	−116.02	1007	2000–2018
345	FPK	Fort Peck	48.31	−105.10	634	2000–2018
346	GWN	Goodwin Creek	34.25	−89.87	98	2000–2018
347	PSU	Penn State	40.72	−77.93	376	2000–2018
348	SXF	Sioux Falls	43.73	−96.62	473	2003–2018
349	TBL	Table Mountain	40.13	−105.24	1689	2000–2018

The first 175 stations are the AmeriFlux sites, followed by 26 AsiaFlux sites (beginning with site code named “MSE”), 57 BSRN sites (beginning with site code named “ALE”), 84 FLUXNET sites (beginning with site code named “AT-Neu”), and 7 SURFRAD sites (beginning with site code named “BND”).

.

Critical quality control procedures were implemented to calculate the daily L_d because the selected networks only provided instantaneous L_d values, except for FLUXNET. The daily mean L_d was integrated from the instantaneous values if the portion of missing instantaneous values was less than 20% in one day. The monthly mean values used for validation were obtained by averaging the effective daily values if the missing daily data reached less than 10 days in one month.

2.1.1. AmeriFlux, AsiaFlux, and FLUXNET Data

FLUXNET [51,52] is a joint regional network that provides continuous measurements of various ecological parameters at five temporal resolutions, including carbon dioxide, water, meteorological data, and radiation data. The FLUXNET2015 dataset contains 1532 site-years of data from 1996 to 2014, of which daily L_d observations are used to build and evaluate L_d estimates over global land surface in this study. The AmeriFlux network [53,54,55] includes 151 sites with more than 100 active sites as of 2012, providing half-hourly or hourly L_d data spanning from 1996 to present. Flux tower sites of the AsiaFlux network [56,57] are spread across various representative climate zones (from humid to arid climates) and land cover types (forest, grass, cropland, and urban area), of which L_d observations have half-hourly or hourly temporal resolutions from 1998 to 2018.

To reduce systematic measurement errors, the data QA/QC checks proposed by Pastorello et al. [58], including single-variable, multi-variable, and specialized checks, are implemented at each site within the three networks. Single-variable checks are aimed at exploring the consistency of one variable in the long and short time series trends. Multi-variable checks focus on the relationship among correlation variables to ascertain discrepant periods. Specialized checks look at common issues in eddy covariance (EC) and meteorological data, such as timestamp shifts or sensor deterioration patterns. The last step for data QA/QC is automatic checks that use specific variable de-spiking routines adapted from Papale et al. [59] to set a range for each variable.

2.1.2. BSRN Data

The baseline surface radiation network (BSRN) was initiated by the world climate research program (WCPR) and aimed to provide accurate observations for validation of satellite radiometry and climate models [60]. The BSRN project has established more than 60 stations globally since January 1992 spanning latitudes ranging from 80°N to 90°S, providing continuous meteorological and radiation data on a minute time scale. By improving its calibration process, the difference between L_d observations from different pyrgeometers only reached 10 W m⁻² in 1995 [61]. Only 6.5% of the L_d data are missing, which indicates that the pyrgeometers within the BSRN maintain high standards [62]. Moreover, the missing data have less influence on L_d because L_d has a small diurnal cycle. Overall, the BSRN L_d observations are relatively accurate and reliable.

2.1.3. SURFRAD Data

The surface radiation budget network (SURFRAD) has provided meteorological and radiation data used for evaluating satellite products and researching climate changes in the United States since 1995. Currently, it is composed of seven stations representing diverse climates with elevations ranging from 98 to 1689 m. It provides long-term and continuous surface radiation measurements with 3 min and 1 min time intervals before and after 2009, respectively. The L_d measured by SURFRAD, with an uncertainty of ±9 W m⁻², covers a wavelength spanning from 4 to 50 μm [63]. The time period of L_d measurements used ranges from 2000 to 2018 in this study.

2.2. Input Data

2.2.1. ERA5 Reanalysis Dataset

ERA5 [64], produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), is the fifth generation reanalysis dataset and a successor of ERA-Interim. It provides complete and consistent hourly temperature, relative humidity, and radiation datasets, in addition to many other atmospheric parameter datasets, with a 25-km spatial resolution from 1979 to near real time. Compared with ERA-Interim [65], ERA5 applied the updated integrated forecast system (IFS) “Cy41r2” 4D-var and produced many new parameters, such as a 100-m wind vector [66]. Many studies have also compared the accuracy of ERA5 and used it to analyze climate change. For example, Wang et al. [66] found that the warm bias of ERA5 2-m air temperature (Ta) is smaller in the warm season and larger in the cold season in relation to the buoy observations over Arctic sea ice. Zhen et al. [67] indicated that the mean relative humidity (RH) of ERA5 displayed a sharp decreasing jump for China during the early 2000s. In this study, the parameters of the ERA5 hourly reanalysis dataset, including the 2-m Ta (°C), the RH at 1000 hPa (%), and the total column water vapor (TCWV, kg m⁻²) from 2000 to 2018, were consolidated into a daily temporal resolution as input data to construct global land L_d (W m⁻²) dataset based on the GBRT method.

2.2.2. GLASS Surface Downward Shortwave Radiation Product

The global land surface satellite (GLASS) daily surface downward shortwave radiation (S_d, W m⁻²) product with a 5-km spatial resolution from 2000 to 2018 was produced from the moderate resolution imaging spectroradiometer top-of-atmosphere (TOA) spectral reflectance on the basis of a direct estimation method [68,69]. First, the TOA reflectance was retrieved using atmospheric radiation transfer simulations under different solar or view geometries. Then, surface shortwave net radiation (S_n) was estimated from the TOA reflectance on the basis of a linear regression relationship between them under different atmospheric conditions and surface properties. Finally, the GLASS daily S_d was produced using daily S_n estimates and surface broadband albedo values. The GLASS daily S_d values obtained an overall RMSE and bias of 32.84 and 3.72 W m⁻², respectively, compared to the ground observations at 525 sites from 2003 to 2005 [68].

2.2.3. Global Multi-Resolution Terrain Elevation Data 2010

The 2010 Global Multi-resolution Terrain Elevation dataset (GMTED2010DEM) [70] is a global continent-wide elevation dataset generated by the U.S. Geological Survey (USGS) and the National Geospatial-Intelligence Agency (NGA). This product contains three spatial resolutions (approximately 250, 500, and 1000 m) aimed at providing generic products for different applications. Carabajal et al. [71] indicated that the GMTED2010DEM products exhibited a great improvement relative to previous elevation data at comparable resolutions. Compared to the global set of the ice, cloud, and land elevation satellite (ICESat) geodetic ground control points, it obtained a positive bias of approximately 3 m. In this study, GMTED2010DEM data with a spatial resolution of approximately 250 m were resampled to a 5-km resolution as input data for estimating L_d to match the generated L_d dataset.

2.3. Exiting Surface Downward Longwave Radiation Datasets

The L_d products used for validation and comparison with the generated L_d dataset contain the clouds and earth’s radiant energy system synoptic (CERES-SYN) edition 4.1 and ERA5 reanalysis datasets. The CERES-SYN product with a 100-km spatial resolution, generated on the basis of the Langley Fu-Liou radiation transfer model [72], provides flux estimates at the TOA and surface, as well as four atmospheric pressure levels (70, 200, 500, and 850 hPa) from 2000 to 2020. Compared with CERES-SYN Edition 3A, the L_d of Edition 4A has been improved due to the improvement of nighttime retrieved cloud properties [73,74]. The ERA5 hourly L_d with a 25-km spatial resolution from 1979 to near real time used the more complicated method proposed by Morcrette [9] to replace the old L_d parametrization [75]. Silber et al. [10] demonstrated that ERA5 underestimated L_d compared with the ground measurements collected from the ARM West Antarctic radiation experiment (AWARE) campaign at McMurdo Station and the West Antarctic Ice Sheet (WAIS) divide. In this study, the daily ERA5 L_d dataset consolidated from the hourly dataset and the CERES-SYN product were compared and used to evaluate the generated L_d dataset from 2000 to 2018.

3. Method

3.1. Gradient Boosting Regression Tree

The gradient boosting regression tree (GBRT) is an ensemble approach that enhances the accuracy of the model by aggregating multiple weak forms of regression and decision trees first proposed by Friedman [76]. The GBRT method is capable of predicting and solving overfitting problems [77]. The core idea of this model is to select the appropriate decision tree function based on the current model and fitting function in order to minimize the loss function. The model produces a strong predictive model by constructing an $M$ amount of different weak classifiers through multiple iterations in order to obtain an accurate prediction rule. Each iteration is to improve the previous results by reducing the residuals of the previous model and establish a new combined model in the gradient direction of the reduced residual [46]. Supposing ${\left\{x_i,y_i\right\}}_{i=1}^N$ is the training dataset, where $x$ represents the predictor variables, $y$ represents the target variable, and $N$ is the number of the training dataset. The GBRT model constructs $M$ different individual decision trees, expressed as ${\left\{h\left(x,{\alpha }_i\right)\right\}}_{i=1}^M$, which can be used to calculate the approximation function of the target variable $f\left(x\right)$ as follows:

$$\{\begin{array}{c}f\left(x\right)={\displaystyle {\displaystyle \sum }_{m=1}^M}{f}_m\left(x\right)={\displaystyle {\displaystyle \sum }_{m=1}^M}{\beta }_{m}h\left(x;{\alpha }_m\right)\\ h\left(x;{\alpha }_m\right)={\displaystyle {\displaystyle \sum }_{j=1}^J}{\gamma }_{jm}I\left(x\in R_{jm}\right), where I=1 if x\in R_{jm}; I=0, otherwise\end{array}$$

(1)

where ${\beta }_m$ and ${\alpha }_m$ are the weight and classifier parameter of each decision tree, respectively. A loss function $L\left(y, f\left(x\right)\right)$ is introduced to describe the accuracy of the model. Each tree partitions the input space into $J$ regions $R_{1m}, R_{2m}, \cdots , R_{jm}$ and each $R_{jm}$ corresponds to a predicted value ${\gamma }_{jm}$. The general process of the GBRT method is shown in Appendix A, Algorithm A1. More details about the GBRT method can be found in Hastie et al. [78] and Ridgeway [79].

The accuracy of the GBRT model which is implemented in the scikit-learn toolbox is mainly affected by its n-estimator, learning rate, max-depth, and subsample parameters. The n-estimator parameter is the maximum number of iterations completed by a weak learner. Larger n-estimators are more likely to lead to overfitting due to a poorer prediction ability with an increasing model complexity. The learning rate parameter is the weight reduction factor of each weak learner, which is usually used together with the n-estimator parameter to determine the fitting effect of the algorithm. The max-depth parameter is the maximum depth of each regression tree, which limits the number of nodes in the tree. The subsample parameter is the proportion of samples used for fitting the base decision tree. Selecting a subsample less than 1 can reduce overfitting but increase the deviation of sample fitting. In this study, the root mean square error (RMSE), mean bias error (MBE), and correlation coefficient (R) between the L_d observations and estimates are used to evaluate the accuracy of the model.

3.2. Model Construction

The daily 2-m air temperature (Ta), relative humidity (RH) at 1000 hPa, total column water vapor (TCWV), surface downward shortwave radiation (S_d), and elevation datasets are selected as predictor variables to estimate the daily surface downward longwave radiation (L_d). The target variable is the daily L_d observations collected at AmeriFlux, AsiaFlux, BSRN, FLUXNET, and SURFRAD from 2000 to 2018. First, the predictor variables were extracted from global datasets corresponding to the ground stations. Then, the dataset of 314 sites was divided into two portions at random: 80% for the training dataset and the remaining 20% for the test dataset. To select the optimal model, 5-fold cross-validation was applied during the training process. The main steps are as follows:

(1)

Calculating daily L_d observations. The daily mean L_d was integrated from the instantaneous values if the missing instantaneous values were less than 20% in one day because the AmeriFlux, AsiaFlux, BSRN, and SURFRAD networks only provide instantaneous L_d values;

(2)

Data preprocessing. After resampling to a 5-km resolution, the ERA5 Ta, ERA5 RH, ERA5 TCWV, GLASS S_d, and GMTED2010DEM elevation datasets were extracted according to the latitude, longitude, and time corresponding to the ground stations;

(3)

Training the GBRT model. By circulating within the range of each parameter displayed in

Table 1. Parameter settings to determine the optimal parameters for the GBRT method.

Parameters	Threshold	Intervals
n-estimator	50–300	50
learning rate	0.1–0.9	0.1
max-depth	4–9	1
subsample	0.2–1	0.1

, the GBRT model where the n-estimator parameter is set to 50, the learning rate is set to 0.1, the max-depth is set to 6, and the subsample parameter of 0.8 was selected as the optimal model to estimate global land L_d, achieving the lowest RMSE and MBE values on the test dataset;

(4)

Implementing the model. The global land L_d was produced on the basis of the trained model using the daily ERA5 Ta, ERA5 RH, ERA5 TCWV, GLASS S_d, and GMTED2010DEM elevation datasets;

(5)

Evaluation of the generated global land L_d dataset. Daily L_d values collected at 35 observation sites were used to validate the generated global land L_d dataset and compare it with the existing L_d datasets. The main flowchart in this study is shown in Figure 2.

In order to investigate the impact of the predictor variables used in the GBRT model on the L_d estimation, the feature importance measures provided by the GBRT method was conducted. As shown in

Table 2. Importance rankings of all predictor variables for L_d estimation.

Predictor Variables	Importance
Total column water vapor (TCWV)	0.78
2-m air temperature (Ta)	0.19
Relative humidity at 1000 hPa (RH)	0.01
Surface downward shortwave radiation (Sd)	0.01
Elevation	0.01

, the importance of the predictor variables of the GBRT model was in the order of the total column water vapor (TCWV), 2-m air temperature (Ta), relative humidity at 1000hPa (RH), surface downward shortwave radiation (S_d), and elevation. The L_d estimates are shown to be more sensitive to the TCWV and Ta than to most of other variables, thus highlighting the importance of taking TCWV and Ta as inputs.

4. Results

4.1. Validation against Ground Measurements

4.1.1. Performance of the Model

After confirming the optimal parameters, 80% and 20% of the extracted dataset collected at 314 stations were used as the training and test datasets, respectively, to train the GBRT model and evaluate the L_d estimates. Figure 3 displays the evaluation results of daily L_d estimates for the training and test datasets against the ground measurements collected at the AmeriFlux, AsiaFlux, BSRN, FLUXNET, and SURFRAD networks from 2000 to 2018. For the training dataset, the root mean square error (RMSE), mean bias error (MBE), and correlation coefficient (R) are 16.73 W m⁻², 0 W m⁻², and 0.96 (p < 0.01), respectively, between the ground observations and L_d estimates on the basis of the GBRT model from 2000 to 2018. Those values are 16.75 W m⁻², 0.05 W m⁻², and 0.96 (p < 0.01) for the test dataset, respectively, which shows a tendency to slightly overestimate L_d. As a whole, the performance of the GBRT model on the test dataset is satisfactory and reliable with an MBE close to zero.

4.1.2. Validation of the Generated L_d Dataset

The L_d observations of 35 sites were used to evaluate the generated L_d dataset collected at the AmeriFlux, AsiaFlux, BSRN, FLUXNET, and SURFRAD networks from March 2000 to December 2018. As shown in Figure 4, the RMSE, MBE, and R values on the daily time scale are 17.78 W m⁻², 0.99 W m⁻², and 0.96 (p < 0.01), respectively, between the ground observations and L_d estimates obtained by the GBRT model. On the monthly time scale, those values are 11.53 W m⁻², 0.68 W m⁻², and 0.98 (p < 0.01), respectively. To further evaluate the performance of the generated L_d dataset, the RMSE, MBE, and R values of the daily L_d estimates at each site were calculated from 2000 to 2018. The minimum and maximum RMSE of the 35 sites are 11.26 and 37.82 W m⁻², respectively. As shown in Figure 5, 24 out of the 35 sites had RMSEs less than 20 W m⁻², and only two sites had RMSEs greater than 30 W m⁻². Overall, 35 sites had absolute MBE values varying from 0.12 to 36.83 W m⁻², and 23 sites had MBEs between −10 and 10 W m⁻². The number of stations with MBE less than −10 W m⁻² and greater than 10 W m⁻² are both six.

4.2. Comparison with Existing L_d Products

To better evaluate the accuracy of the generated L_d dataset, the valuation result against the 35 sites from 2000 to 2018 was compared with the CERES-SYN and ERA5 products. The generated L_d and ERA5 products were resampled to a 100-km resolution using the nearest neighbor interpolation method to match the CERES-SYN product. As shown in Figure 6, the RMSE and MBE are 17.94 and 0.25 W m⁻², 18.81 and 1.76 W m⁻², 18.52 and −2.09 W m⁻², respectively, for the daily generated, CERES-SYN, and ERA5 L_d datasets. CERES-SYN and ERA5 L_d show overestimated and underestimated tends on the daily time scale, respectively. Relatively speaking, the overestimated trend of the generated L_d dataset with an MBE of 0.25 W m⁻² is slight. In addition, the RMSE of the ERA5 daily L_d dataset is less than that of the CERES-SYN product, and it can be concluded that the ERA5 L_d product over land is more accurate than that of the CERES-SYN on the daily time scale. This is consistent with the conclusion of Tang et al. [11] that the ERA5 L_d product over land surface has a higher accuracy on average than the CERES-SYN on the hourly, daily, and monthly time scales but has a worse accuracy than the CERES-SYN dataset over ocean surface. On the monthly time scale, the RMSE and MBE are 11.75 and 0.18 W m⁻², 13.55 and 1.63 W m⁻², 12.20 and −2.69 W m⁻², respectively, for the generated, CERES-SYN, and ERA5 L_d datasets. It can be concluded that the generated L_d dataset based on the GBRT model performed best on both daily and monthly scales. To further compare the performance of the three daily L_d datasets, the RMSE, MBE, and R values at each site were calculated from 2000 to 2018. The RMSE of the 35 sites varied from 11.21 to 31.90 W m⁻², 9.09 to 41.99 W m⁻², 8.68 to 35.51 W m⁻², respectively, for the daily generated, CERES-SYN, and ERA5 L_d datasets. As shown in Figure 7, there are 28, 28, and 25 sites with RMSEs less than 25 W m⁻² for the daily generated, CERES-SYN, and ERA5 L_d datasets. Only 3, 3, and 2 out of 35 sites had RMSEs greater than 30 W m⁻² for the three daily L_d datasets, respectively. These three daily L_d datasets have 23, 19, and 24 sites with MBEs between −10 and 10 W m⁻², respectively. However, the daily CERES-SYN L_d product obtained 10 sites with MBEs greater than 10 W m⁻², compared with 6 for the generated L_d dataset and 3 for the ERA5 L_d retrieval.

4.3. Spatial and Temporal Analysis of L_d

4.3.1. Spatial Distribution

The multiyear seasonal and annual mean values of the generated L_d dataset from 2003 to 2018 (i.e., not from 2000 to 2018) were calculated due to the absence of daily L_d values from 2000 to 2002. The CERES-SYN and ERA5 L_d products were resampled to a 5-km resolution by the bilinear interpolation method for comparison with the generated L_d dataset. The spatial distributions of the multiyear seasonal and annual mean L_d estimations over the global land surface from 2003 to 2018 are displayed in Figure 8 and Figure 9, respectively. The highest multiyear seasonal mean L_d value is 333.21 W m⁻² in Northern hemisphere summer (June, July, and August), followed by 311.94 W m⁻² in Northern hemisphere autumn (September, October, and November), and the lowest value is 286.09 W m⁻² in Northern hemisphere winter (December, January, and February). The seasonal variation in L_d is closely related to the annual solar zenith cycle and the maximum sunshine duration. After the winter solstice, the direct sun point moves northward from the Tropic of Capricorn, causing changes in the global heat distribution, which increases the overall L_d value in the Northern hemisphere. Overall, the multiyear annual mean value of the generated L_d dataset is 308.76 W m⁻², which is greater than the ERA5 value of 306.92 W m⁻² and less than the CERES-SYN value of 313.83 W m⁻² from 2003 to 2018. The spatial distribution of L_d not only shows significant latitudinal dependencies in which the mean L_d value decreases with increasing latitude but also relates to the surface elevation and regional climate. The mean L_d values estimated over the Andes and Tibetan Plateau are comparatively and obviously low due to their high elevation with a low cloud coverage, thin air and readily lost heat. The mean L_d values of Antarctica and Greenland are always lowest owing to the perennial snow cover and frigid climate. Apparently, the generated L_d value is lower than the CERES-SYN value and higher than the ERA5 value. The lowest and highest differences between the generated L_d and the CERES-SYN product are −81.44 and 60.56 W m⁻², respectively; and the values between the generated L_d and the ERA5 product are −46.17 and 58.83 W m⁻², respectively. The generated L_d value is significantly lower than the CERES-SYN in the Tibetan Plateau, Andes Mountains, and Antarctica, and is significantly higher than it in a small area of the northern Amazon Rainforest and eastern Indonesia. Compared with the CERES-SYN dataset, the difference between the generated L_d dataset and the ERA5 product is evenly distributed with no obvious high and low values over the global land surface.

The multiyear annual mean L_d values of the generated dataset, ERA5 retrieval, and CERES-SYN product are consistent with the evaluation results against the ground measurements that ERA5 and CERES-SYN tend to underestimate and overestimate L_d value, respectively. However, there is still much debate about the specific multiyear annual mean L_d value over the global land surface. The uncertainty of the global land mean L_d estimation is difficult to quantify, and different periods may influence the estimated values. Ma et al. [80] summarized that multiyear annual mean L_d values over the global land surface varied between 287.35 and 316.62 W m⁻² for 44 general circulation models (GCM) in the coupled model intercomparison project phase 5 (CMIP5) from 1990 to 2005, and its median was 304.59 W m⁻². Wang et al. [81] calculated that the annual mean L_d values over the global land surface of the GEWEX-SRB, MERRA, and CERES-SRB datasets were 308, 295, and 307 W m⁻² from 2001 to 2007, 2001 to 2010, and 2003 to 2010, respectively, and estimated that the best L_d estimate was 307±3 W m⁻² over the global land surface from 2003 to 2010 based on reference studies and evaluation results compared against the ground measurements. The multiyear annual mean value of the generated L_d dataset over the global land surface is 308.76 W m⁻² from 2003 to 2018 which is consistent with these results.

4.3.2. Time Series and Long-Term Trend

To study the temporal variations of the generated L_d dataset, we calculated the monthly and annual mean L_d from 2003 to 2018, as shown in Figure 10. We analyzed whether the interannual changes in the generated L_d dataset were reliable by comparing with the ERA5 and CERES-SYN L_d datasets. Overall, ERA5 has a relatively lower value, and CERES-SYN shows a larger value for the multiyear monthly mean L_d. The multiyear monthly mean L_d of the ERA5, generated dataset, and CERES-SYN are all lowest in January, with values of 280.02, 284.22, and 284.49 W m⁻², respectively; they are all largest in July, with values of 336.12, 336.24, 343.94 W m⁻², respectively. The multiyear monthly mean L_d values of the three datasets all increase from January to July and decrease from July to December in connection with the revolution of the earth around the sun resulting in more total solar radiation in Northern hemisphere summer than in Northern hemisphere winter. Compared with ERA5 and CERES-SYN, the boxed part of the box-plot (Figure 10a) of the generated L_d dataset is relatively compact, indicating that its monthly mean L_d values in different years are concentrated. Similar to the monthly mean L_d, the annual mean L_d values of ERA5 and CERES-SYN are lower and higher than the generated L_d dataset, respectively, in the same year. The annual mean L_d values ranged from 304.93 to 309.92 W m⁻², 306.94 to 311.99 W m⁻², and 311.88 to 316.29 W m⁻² from 2003 to 2018, respectively, for the ERA5 retrieval, the generated L_d dataset, the CERES-SYN product. The three datasets all obtained the lowest and largest annual mean L_d values in 2008 and 2016, respectively.

As displayed in Figure 10c, before 2015, the anomalies of the annual mean L_d values are negative for the generated and ERA5 L_d datasets, except for 2005 and 2010, which implies that the annual mean L_d values for this period are below the multiyear average over 16 years. In addition, the anomalies of annual mean L_d values are also more than zero for the CERES-SYN product in 2003. Moreover, the CERES-SYN L_d product had a smaller growth trend of 0.8 W m⁻² per decade (p = 0.20) from 2003 to 2018, but the growth trend was not significant. Overall, the temporal variation and trend of the generated L_d dataset are more consistent with the ERA5 product, and the annual mean L_d values display a gradual increasing trend from 2003 to 2018. Ma et al. [80] concluded that the trend of the annual mean L_d over the global land surface for 44 CMIP5 GCMs varied from 0.69 to 2.86 W m⁻² per decade (p < 0.01) during the time period of 1970–2005, and its median value is 1.86 W m⁻² per decade. Therefore, it is reliable for the generated and ERA5 L_d datasets, with trends of 1.8 (p < 0.01) and 1.9 (p < 0.01) W m⁻² per decade over the global land surface, respectively.

4.3.3. Relationships between the Long-Term L_d and the Key Factors

Previous studies indicated that the accuracy of L_d estimation mainly depends on the reliability of the air temperature, precipitable water vapor, cloud, and elevation data retrieved from the reanalysis and satellite products. In view of the small variations in elevation and cloud cover over the long time series, the trend of L_d estimation is mainly influenced by air temperature and water vapor pressure. Therefore, we calculated the anomalies of the 2-m air temperature (Ta, °C) and water vapor pressure (e, hPa) datasets based on the ERA5 hourly products from 2003 to 2018. e can be calculated with Ta using the following equations based on the Ta and relative humidity at 1000 hPa (RH) derived from the ERA5 products.

$$\mathrm{RH}=\frac{e}{e_s}\times 100\%$$

(2)

$$e_s=6.11\exp \left(\frac{L_v}{R_v}\left(\frac{1}{273.15}-\frac{1}{\mathrm{Ta}+273.15}\right)\right)$$

(3)

where e and e_s are the water vapor pressure and saturation water vapor pressure, respectively.

Figure 11 presents the temporal variation in the annual mean anomalies for the generated L_d estimation, Ta and e from 2003 to 2018. The annual mean values ranged from 9.12 to 10.22 °C, and 7.82 to 8.26 hPa from 2003 to 2018 for Ta and e, respectively. Before 2015, the annual mean anomalies were negative for Ta and e, excluding 2005 and 2010, which was similar to the L_d estimation. In addition, the anomalies of annual mean Ta values were also greater than zero in 2007. The L_d increases with the increase in Ta and e. The increasing rates are 0.3 °C per decade (p < 0.01), 0.1 hPa per decade (p < 0.05), and 1.8 W m⁻² per decade (p < 0.01), respectively, for Ta, e, and L_d from 2003 to 2018. Overall, Ta and e positively influence L_d with correlation coefficients of 0.96 (p < 0.01) and 0.97 (p < 0.01), respectively. The strong absorption and re-emission of radiation by water vapor molecules result in a high correlation between e and L_d. However, the influence of temperature on L_d relies on the dependence of the outgoing longwave radiation on the absolute temperature of the Earth. In addition, the spatial distributions of the annual mean values of Ta and e from 2003 to 2018 are shown in Figure 12. The minimum and maximum annual mean values are −53.35 and 34.12 °C, and 0.05 and 32.85 hPa, respectively, for Ta and e. Their distribution characteristics are similar to that of the generated L_d dataset that its spatial distribution not only shows notable latitudinal dependencies as the annual mean values decrease with increasing latitudes but it also relates to the surface elevation and regional climate. The annual mean Ta and e values on the Andes and Tibetan Plateau are comparatively and obviously low due to their high elevations. The annual mean Ta values of Antarctica and Greenland are always the lowest due to their perennial snow coverage and frigid climates. The annual mean e values are relatively low, less than 21.72 hPa at middle to high latitudes. The spatial distribution of R between the generated L_d estimation and Ta and e from 2003 to 2018 is also drawn, as shown in Figure 13. Only significant pixels where p values are less than 0.05 appeared. The R values ranged from 0.50 and −0.67 to 1 for Ta and e, respectively. There was a positive correlation between the generated L_d and Ta in the region where the R passed the significant test. For the e, there are few pixels with R value less than 0, which even cannot be shown up on the map. Except for values less than 0, the minimum value of R between the generated L_d and e is also 0.50. The R values between annual mean L_d estimates and Ta and e failed the significant test mainly occurred in the Andes Mountains, Brazilian Plateau, Tibet Plateau, Australia, Southern Africa, and southern North America. This may be due to the influences of clouds, elevation controls, and carbon dioxide emissions [29,30,41,82], that play a dominant role in these regions. The possible reasons need to be further explored.

5. Discussion

5.1. Shortcomings of the GBRT Model

The gradient boosting regression tree (GBRT) method has advantages in forecasting and solving overfitting problems [76]. The evaluation results demonstrated that the generated L_d dataset based on the GBRT method performed better at selected stations than the ERA5 and CERES-SYN products on daily and monthly time scales. However, there are still some disadvantages in the machine learning methods for radiation estimation represented in the GBRT method [45,77,83]. Alizamir et al. [77] utilized six different machine learning models to estimate solar radiation from two stations of two different locations, and found that the six models all tend to overestimate L_d for low values of it and to underestimate L_d for high values of it. Fan et al. [83] also exposed similar problem in using support vector machine and extreme gradient boosting methods to predict daily solar radiation in China. With reference to Figure 3, Figure 4 and Figure 6, it is evident that the GBRT method for L_d estimation also overestimate L_d at low values and underestimate L_d at high values. Since the departures of slope from 1 and intercept from 0 for fitting linear regression equations can measure the degree of deviation, the linear regression equations were fitted between the ground measurements of 35 stations and the generated, ERA5, and CERES-SYN L_d datasets on daily and monthly time scales, as listed in

Table 3. The fitted linear regression equations for the generated, ERA5, and CERES-SYN L_d datasets on both daily and monthly time scales. Where the x and y represent the ground measurements of L_d and the L_d estimates, respectively.

Time Scale	Dataset	Fitted Linear Regression Equation
Daily time scale	Ld estimation	$y=0.91\ast x+27.99$*
	ERA5 Ld	$y=0.97\ast x+6.39$*
	CERES-SYN Ld	$y=0.96\ast x+13.38$*
Monthly time scale	Ld estimation	$y=0.94\ast x+19.06$*
	ERA5 Ld	$y=0.99\ast x+1.64$*
	CERES-SYN Ld	$y=x+0.59$*

* The coefficient of the fitted linear regression equation passed the significance test (p < 0.01).

. Compared to those two L_d products, the fitted linear regression equations of the generated L_d has a smaller slope and a greater intercept on both daily and monthly time scales, which indicates that the fitted line deviates more from the 1:1 line and that the GBRT method underestimates L_d for high values and overestimates it for low values. Other machine learning methods, including support vector regressions, multivariate adaptive regression splines, and artificial neural networks used to estimate L_d, also show the same problem [49]. On the other hand, the GBRT method makes predictions by learning rules from many sample data, so it has higher requirements for the accuracy and quantity of its training datasets. However, the ground measurements of L_d used as the target variable exhibit missing values and deviations, although the obviously incorrect data have been removed, which may be limit the accuracy of the GBRT method. Finally, the learning and training process of the GBRT method is a black box whose processes are not known and may not be effective [45].

5.2. Accuracy and Completeness of Input Datasets and Ground Measurements

Because L_d was estimated based on the relationships between its ground observations and input variables, including the 2-m air temperature (Ta), relative humidity (RH) at 1000 hPa, total column water vapor (TCWV), surface downward shortwave radiation (S_d), and elevation, the accuracy and completeness of the input datasets and ground measurements are vital. However, ground observations exist measurement errors and problem of spatial representativeness, which are potential sources of errors in L_d estimation. A larger part of measurements errors is caused by systematic deviations and calibration process differences. Ohmura et al. [60] demonstrated the accuracy of L_d observations in the baseline surface radiation network improved from 30 W m⁻² in 1999 to 10 W m⁻² in 1995 due to improvement of the calibration process. Currently, although pyrgeometers for L_d measurement are regularly maintained and calibrated, there is still a lack of recognized world reference calibration standard [60,84]. The different calibration methods of different observation networks can lead to inconsistencies of L_d measurements at the close positions, which also brings uncertainties of ground measurements [81]. The spatial representativeness plays an important role in the surface radiation retrieval and validation [85,86,87,88]. Jiang et al. [87] indicated the accuracy of S_d retrieval can be improved, that maximum improvement of root mean square error is up to 9%, after considering the scale information. In this study, we only compared the accuracies of the generated, ERA5, and CERES-SYN L_d datasets at 100-km spatial resolution but did not examine the representativeness of surface observation points, which maybe lead to the uncertainty of compared result.

On the other hand, the completeness of input datasets limits the continuity of the generated L_d dataset. For example, the generated daily L_d dataset was discontinuous before 2003 due to the data missing of the global land surface satellite (GLASS) daily S_d product which was produced by using moderate resolution imaging spectroradiometer (MODIS) top of atmosphere (TOA) reflectance data. Compared with the previous methods of S_d retrieval, however, GLASS S_d had a higher spatial resolution of 5 km and was directly estimated using TOA reflectance without the need for cloud and aerosol data, which contributes to a better ability to demonstrate temporal variations in S_d over a long time period. Moreover, the ERA5 and elevation datasets were resampled to a 5-km spatial resolution matching with S_d, which can also introduce uncertainty into the data. In summary, the L_d estimates will be more accurate if the accuracy and completeness of the input datasets and ground measurements are improved.

6. Conclusions

It is of great importance for studying the Earth’s surface radiation budget and energy balance to construct a long-term surface downward longwave radiation (L_d, 4–100 μm) dataset worldwide. This study generated a daily L_d dataset with a 5-km spatial resolution over the global land surface utilizing the gradient boosting regression tree (GBRT) method with 2-m air temperature (Ta), relative humidity (RH) at 1000 hPa, total column water vapor (TCWV), surface downward shortwave radiation (S_d), and elevation datasets from 2000 to 2018. The L_d observations of 349 stations collected at the AmeriFlux, AsiaFlux, baseline surface radiation network (BSRN), surface radiation budget network (SURFRAD), and FLUXNET networks were randomly divided into 90% (314 sites), as the target variable to build the model, and 10% (35 sites), as the evaluation dataset to independently validate the L_d estimates. First, the predictor variables were extracted from the global datasets according to the latitude, longitude, and time corresponding to the ground stations. Then, the dataset of 314 sites was further divided into two portions at random to train the GBRT model: 80% for the training dataset and the remaining 20% for the test dataset. Then, the daily L_d observations collected at 35 stations were used to validate the generated global land L_d dataset.

The evaluation results showed that the root mean square error (RMSE), mean bias error (MBE), and correlation coefficient (R) values on the daily time scale were 17.78 W m⁻², 0.99 W m⁻², and 0.96 (p < 0.01), respectively, between the L_d estimates with a 5-km spatial resolution and the ground measurements. On the monthly time scale, those values are 11.53 W m⁻², 0.68 W m⁻², and 0.98 (p < 0.01), respectively. At a 100-km spatial resolution, the performance of the generated L_d dataset is better than that of ERA5 and CERES-SYN. On the daily time scale, the RMSE and MBE are 17.94 and 0.25 W m⁻², 18.81 and 1.76 W m⁻², 18.52 and −2.09 W m⁻², respectively, for the generated, CERES-SYN, and ERA5 L_d datasets. The multiyear seasonal and annual mean values of the generated L_d dataset from 2003 to 2018 were calculated due to the absence of daily L_d from 2000 to 2002. In terms of their temporal variation, the multiyear monthly mean L_d values of the three datasets increase from January to July and decrease from July to December in connection with the revolution of the earth around the sun resulting in more total solar radiation in Northern hemisphere summer than in Northern hemisphere winter. Overall, the temporal variation and trend of the generated L_d dataset are more consistent with the ERA5 product that the annual mean L_d values display a gradual increasing trend from 2003 to 2018. The spatial distribution of L_d not only shows a notable latitudinal dependency in which the mean L_d value decreases with increasing latitudes but also relates to the surface elevation and regional climate. In addition, L_d is positively affected by the 2-m air temperature and water vapor pressure with R values of 0.96 (p < 0.01) and 0.97 (p < 0.01), respectively.

Overall, the generated L_d dataset has a higher spatial resolution and accuracy, contributing to knowledge of the surface radiation budget and energy balance of the Earth.