The Land Surface Temperature product is the 6.1 version, developed by NASA's Land Processes Distributed Active Archive Center (LP DAAC). The data is extracted at a 0.05° latitude/longitude grid and provides monthly LST and emissivity data derived from MODIS, carried aboard the Terra (Wan et al., 2021) satellite. regridded to 1° × 1° grid size for discussion and analysis. NASA's Terra is a sun-synchronous, polar orbiting satellite in a morning orbit, with equator crossing times of approximately 10:30 am and 10:30 pm (local time, LT) in descending and ascending modes respectively. In this work, we focus on the Terra satellite because of its close crossing time to the other datasets listed hereafter. Terra (and Aqua) have been drifting to an earlier equatorial crossing time since 2020. NORAD (North American Aerospace Defense Command) data (not shown here) suggests that by the end of 2022, the equatorial crossing time was around 10:11 pm LT, for a total of 19 min drift. The ∼ 19 min drift remains small relative to the monthly sampling used in this study and should not introduce overpass-condition changes capable of biasing the comparisons between the different datasets.

To complement our analysis, we use the ESA LST CCI Monthly Multisensor Infrared (IR) Low Earth Orbit Land Surface Temperature Level 3 Supercollated (L3S) global product, version 3.00 (Ghent et al., 2025), which has been online since August 2025.

Daytime and nighttime temperatures are provided in separate files, corresponding to local solar times of 10:00 am and 10:00 pm, respectively. The instruments contributing to the time series used in this study are AATSR (Advanced Along-Track Scanning Radiometer, until March 2012), MODIS Terra (from April 2012 to November 2018), and SLSTR-B (the Sea and Land Surface Temperature Radiometer on Sentinel 3B, starting in December 2018). For consistency, a common algorithm is applied for LST retrieval across all instruments. In addition, an adjustment is made to account for the half-hour difference in the satellites' equatorial crossing times; this correction is computed in brightness temperature space using radiative transfer simulations (Ghent et al., 2025). Data are retrieved at a 0.01° latitude/longitude grid, that we regridded to 1° × 1° grid size for discussion and analysis.

An important characteristic of the ESA LST CCI that distinguishes it from the NASA MODIS LST product is the restriction of retrievals to view zenith angles (VZA) of ±22°. This restriction is intentional: the full LST CCI data record also includes data from ATSR-2 and AATSR, which have narrower swaths than MODIS or SLSTR, and limiting all sensors to ±22° ensures geometric consistency across the entire multi-decadal record. In contrast, the NASA MODIS LST product uses the full MODIS swath of ±60° VZA. It has been shown that MODIS LSTs are systematically depressed at the edges of the swath due to increased atmospheric path length and angular emissivity effects at high viewing angles (Ermida et al., 2019; Wan, 2014). By averaging over a much wider angular range that includes these colder off-nadir retrievals, the MODIS global mean LST is expected to be lower than the near-nadir-only LST CCI product. This viewing geometry difference is the primary physical explanation for the systematic warm offset of LST CCI relative to MODIS (discussed in Fig. 1), and does not reflect a calibration error.

To extend the ocean surface temperature comparison with an independent satellite-based dataset, we include the ESA SST CCI and C3S Reprocessed Sea Surface and Sea Ice Temperature product, distributed through the Copernicus Marine Service (Høyer et al., 2014; Høyer and She, 2007; Nielsen-Englyst et al., 2023). We use the variable “st”, which corresponds to the analysed sea (at 20 cm depth) and ice surface (surface skin) temperature. This Level-4 (L4) product provides gap-free daily analyses at a native horizontal resolution of 0.05° × 0.05°.

The product is generated using the Danish Meteorological Institute Optimal Interpolation (DMIOI) system, which combines input from multiple satellite sources (Embury et al., 2024); and IST retrievals from the AASTI (Arctic and Antarctic Surface Temperatures from thermal Infrared satellite radiometers) and C3S IST Climate Data Record (CDR)/Interim CDR v.1. The inclusion of IST, constrained by a multi-source sea-ice concentration composite, makes this product particularly suited for representing surface temperatures in polar regions, the sea ice, and the marginal ice zone – areas where other SST products are typically undefined or unreliable. The product is extracted over the study period of 2008–2022, which we regrid to a 1° × 1° grid for consistency with the other datasets used in this study. In this work, this dataset is the only one that is all-sky, and not separated between day and night. We refer to it as ESA CCI/C3S SST and it serves as such as a reference to the other SST datasets.

We use the Tskin variable from the ERA5 reanalysis (Hersbach et al., 2020), produced by ECMWF, which is framed within the Copernicus Climate Change Service of the European Commission. ERA5 datasets are at 0.25° × 0.25° resolution (native horizontal resolution of ERA5 is ∼ 31 km), that are regridded to a 1° × 1° grid size for the study.

Hourly averages are computed for each month of the year and are then converted to local time by calculating the time zone offset based on the longitude (15° = 1 h). Our analysis does not account for daylight saving time. This is consistent with satellite local-solar-time definitions, as polar-orbiting instruments such as IASI maintain fixed solar crossing times (e.g., 09:30/21:30 LT), independent of civil time zones.

ERA5 should not be treated as an independent ground truth, but rather as a physically consistent reference that is subject to its own model-dependent uncertainties. Over land, ERA5 Tskin is a prognostic variable computed within the HTESSEL (Hydrology-Tiled ECMWF Scheme for Surface Exchanges over Land) land surface scheme (Balsamo et al., 2009), as part of the Integrated Forecast System (IFS). It is integrated forward in time by solving the surface energy balance at each model timestep, and is therefore sensitive to model parameterisations of soil thermal properties, vegetation, and snow. Over ocean, ERA5 Tskin is not prognostic but is instead derived from a prescribed SST analysis with a diagnostic cool-skin correction applied (Hirahara et al., 2016; Zeng and Beljaars, 2005). ERA5 does not directly assimilate Tskin, but does assimilate SST analyses over ocean and infrared radiances over both land and sea (Hersbach et al., 2020), which carry indirect surface temperature information.

IASI (Clerbaux et al., 2009) is a Fourier transform spectrometer and data are available from three instruments onboard of Metop-A, -B and -C satellites launched in 2006 (end of life in October 2021), 2012, 2018, respectively. Each IASI instrument provides more than 1.2 million radiance spectra per day with a footprint on the ground of 12 km diameter at nadir, and a local crossing time of 09:30 am and 09:30 pm LT. IASI has exceptional spectral and radiometric stability making it a reference instrument (Kilymis et al., 2025). IASI Tskin data are officially distributed by EUMETSAT and is derived from IASI radiances but also relies on two microwave instruments onboard Metop satellites – the Microwave Humidity Sounder (MHS) and the Advanced Microwave Sounding Unit-A (AMSU-A) – particularly for cloudy scenes. Several updates have been made to the processing algorithm over time (Bouillon et al., 2020), which has been relatively stable since 2016, making the product operational but not homogeneous over time. To this end, EUMETSAT provided an all-sky climate data record (IASI-CDR) Tskin product. In this work we use the latest CDR version, 1.1, which is a homogeneously reprocessed dataset from the surface to the top of the atmosphere (EUMETSAT, 2022). The reprocessed temperatures were computed with a piecewise linear regression cube (PWLR³) algorithm, using all IASI observations as input (clear and cloudy scenes). It is a fast and accurate statistical retrieval scheme that exploits the synergy between microwave (MHS, and AMSU) and infrared measurements. PWLR³ is a machine learning algorithm trained with real satellite observations paired with the best correlative representation of the Earth system. The CDR includes four surface parameters (surface pressure, surface air temperature, surface air dew point temperature, and surface skin temperature) as well as the atmospheric profiles of temperature and humidity. Only surface skin temperature is used in this work. The training set used for the algorithm in this CDR is composed of three variables from the European Reanalysis (ERA5): temperature, humidity and ozone, provided at 137 model levels (Hersbach et al., 2020). It uses 96 d of data, i.e., 4 d (1st, 8th, 15th and 22nd) of each month for 2 years in 2015 and 2017. In total, the training set is made up of more than 120 million IASI fields of view. The algorithm defines different observation classes based on the IASI observations and auxiliary information (e.g. surface type and elevation). Then, for each observation class, a linear regression is performed from the IASI/AMSU-A/MHS observation to retrieve the required geophysical parameters.

The IASI Neural Network (IASI NN) Tskin product used in this study is a clear-sky retrieval based on an information-content channel selection and a neural-network architecture. It is an updated version of the product validated by Safieddine et al. (2020). Surface temperature is derived using 87 window-region channels (750–950 cm⁻¹), reduced from 100 in the first version by removing channels sensitive to carbon dioxide (CO₂) to improve long-term stability. The list of channels is provided in Appendix A. A second key update is the use of the EUMETSAT Near Real Time (NRT) Tskin product from 2021 as the training target, replacing ERA5, as was done in Safieddine et al. (2020). This allows the NN to emulate an operational optimal-estimation surface temperature directly derived from IASI and ensures consistency over the entire mission. The NRT product is based on optimal estimation and over the years, EUMETSAT has performed several updates on the real-time processing of both radiances and temperatures, making the time series non-homogeneous, until around 2017 (Bouillon et al., 2020). For this reason, the year 2021 is used as a trustworthy year for the training target.

The datasets used in this study are not fully independent of one another, which is relevant when interpreting intercomparison results. ERA5 assimilates IASI radiances for atmospheric profile retrieval (Hersbach et al., 2020), which partially constrains its land surface scheme and thus its Tskin. IASI-CDR in turn uses ERA5 atmospheric profiles as training input, introducing a dependency on ERA5. IASI-NN uses the EUMETSAT NRT product as its training target, making it more independent than IASI-CDR on ERA5; however, it incorporates MODIS emissivities as auxiliary input, linking it partially to the MODIS instrument which retrieves LST directly from infrared radiances (and is effectively independent of ERA5 Tskin). Finally, LST CCI, although also derived from infrared radiances, was intercalibrated against IASI (Ghent et al., 2025), introducing a degree of dependency between these two datasets. The ESA CCI/C3S SST and IST product is derived from an independent suite of satellite radiometers and is anchored to in-situ drifting-buoy measurements (Embury et al., 2024), making it largely independent of both the IASI and ERA5 datasets with SST products used here.

Over sea-ice-covered regions, the definition and measurement of Tskin becomes more complex. Infrared satellite retrievals such as IASI and MODIS are sensitive to the radiometric skin temperature of the ice surface, but their accuracy depends on cloud screening and surface emissivity characterisation, both of which are particularly challenging in polar regions. ERA5 treats sea ice explicitly through a dedicated surface scheme. In this study, sea-ice-covered areas are not analysed separately and are included within SST, and the inter-dataset biases observed at high latitudes and discussed later are partly attributable to these differences in sea-ice treatment across datasets.

To ensure a physically consistent intercomparison, all datasets are evaluated under clear-sky conditions. Each dataset is filtered using the most appropriate cloud screening method. For IASI-NN, clear-sky retrievals are already guaranteed by the cloud mask of Whitburn et al. (2022), which is applied at the pixel level prior to retrieval. The same cloud mask is applied to IASI-CDR, retaining only those IASI-CDR retrievals collocated in space and time with IASI-NN clear-sky pixels. Since both products share the same instrument, orbital swath, and overpass times, this constitutes a physically consistent filter that ensures both IASI products are evaluated over identical scenes.

For ERA5, we apply a threshold on the total cloud cover (tcc) variable, ERA5's native cloud diagnostic, retaining only grid points and time steps where tcc < 0.2 (20 %). This clear-sky criterion is consistent with the approach used in comparable studies (Ermida et al., 2019; Ermida and Trigo, 2022), which also demonstrate through sensitivity analysis that varying the cloud fraction threshold between 1 % and 30 % changes the derived clear-sky bias by less than 0.5 K for 85 % of grid points. Moreover, the ERA5 Tskin trends derived from monthly cloud-filtered data are shown to be robust across a wide range of cloud cover thresholds (Wang et al., 2022).

MODIS Terra LST (v6.1) and ESA LST CCI (v3.00) are clear-sky products by construction, as their retrieval algorithms only produce valid retrievals under cloud-free conditions; no additional cloud filtering is therefore applied to these datasets.

The ESA SST CCI/C3S SST and IST product combines infrared clear-sky retrievals with all-sky passive microwave observations through optimal interpolation, and is therefore not cloud-filtered. Applying a cloud mask, a posteriori, using ERA5 monthly cloud cover fraction, for example, would be the only feasible approach to approximate clear-sky conditions, but this would introduce an ERA5-dependent filtering that is inconsistent with the cloud screening applied to IASI and MODIS, which is based on their own instrument-level detection. We therefore present the ESA CCI/C3S SST and IST product as an all-sky all-day reference.

We show in Fig. 1, the global averages of LST (upper panels), and the deseasonalised anomaly (lower panels) from the 5 LST datasets used in this study (MODIS, CCI, ERA5, IASI CDR, IASI NN), during the day (left) and the night (right). For the Tskin, i.e., land + sea datasets (ERA5, IASI CDR et IASI NN), a sea mask is applied based on ERA5 (land/sea mask).

During daytime, LST CCI is systematically warmer than the other datasets in the global mean (Fig. 1, upper-left), yielding the highest LST values throughout most of the record. The deseasonalised anomalies display very good agreement among the different datasets, except for LST CCI. The shift from AATSR to MODIS/Terra in April 2012 is clearly visible in the anomaly time series, as is another shift in late 2018 (nighttime), corresponding to the introduction of SLSTR-B into the CCI dataset. This version of the CCI product has not yet been validated. Here we show that the addition of new sensors appears to affect the homogeneity of the LST time series.

At night, MODIS provides the lowest LST values. MODIS has the latest overpass time among the products considered, so we would expect that Tskin at 10:30 pm LT (Terra) to be lower than those at 09:30 pm LT (IASI) or at 10:00 pm LT (ERA5 and LST CCI). However, the magnitude of the MODIS cold bias cannot be explained solely by differences in equator-crossing times. To investigate this further, Fig. 2 presents the spatial bias of MODIS with respect to the four other datasets. A pronounced nighttime cold bias is evident across all comparisons, including LST CCI, which includes MODIS data (but a distinct retrieval algorithm, see Sect. 2). Tan et al. (2021) validated MODIS LST against ground measurements and found a negative nighttime bias at most stations, with an average of -1.58 K. In high-latitude regions, cold biases have also been reported in the Arctic (Muster et al., 2015) and in Antarctica (Zhai et al., 2024), often linked to cloud-screening issues that affect the retrievals. Surface heterogeneity – including snow/ice, permafrost, and arid regions – can further influence MODIS LST (Hachem et al., 2012), though these challenges extend to other satellite products as well. In urban areas, nighttime MODIS LST retrievals have been shown to exhibit non-negligible errors (Yoo et al., 2022).

Figure 2 highlights the widespread MODIS negative (cold) bias over land at night. This suggests that the MODIS V6.1 LST product used here may exhibit a generalized nighttime cold bias that warrants further investigation. The daytime bias is smaller in magnitude, especially with LST CCI, and more spatially variable. When comparing MODIS to ERA5, the longitude “stripes” appears because of the local time conversion method (see Sect. 2.3). IASI-NN and MODIS agrees more in arid and ice-covered region likely because our neural-network product incorporates emissivity information derived (partly) from MODIS. Pronounced discrepancies between MODIS and ERA5 (as well as IASI-CDR trained with ERA5) over the Tibetan Plateau are consistent with known limitations of reanalyses in high-altitude and complex terrains (Khadka et al., 2022; Liu et al., 2024).

In Fig. S2, we show the RMSE for the same pairs of datasets. The RMSE is generally between 1 and 3 K over most land surfaces during both day and night, with higher values exceeding 4–5 K over arid regions and at high latitudes. These elevated RMSEs reflect local surface heterogeneity and variable emissivity conditions rather than a systematic global offset. The consistency between the spatial patterns of bias (Fig. 2) and RMSE (Fig. S2) indicates that the global mean bias values (reported in Fig. 3) are not artefacts of spatial cancellation, but represent a genuine, if spatially non-uniform, systematic difference between the datasets.

We then show and compute the global averaged bias and root mean square errors (RMSE) over land areas, and between the different datasets in Fig. 3. Both metrics were computed as spatially weighted averages using cos(latitude) weighting to account for the varying grid-cell area with latitude.

The datasets show good agreement during the day, with biases between the different products lower than 1 K (except for LST CCI, with biases between -2.5 and 2.7). RMSE values are between 2 and 3 for the datasets except for LST CCI. RMSE values are the highest for the LST CCI dataset during the day, representing the overall lower accuracy of this dataset. During the night, both the biases and RMSE are higher for MODIS with respect to the other datasets, as it was shown in Figs. 1 and 2. The IASI-NN product shows biases below 1 K when compared with MODIS, ERA5, and the IASI-CDR dataset during the day, and has an average bias of 1.94 K relative to LST CCI. At night, IASI-NN agrees well with the other products (< 1.7 K), except for MODIS, as shown and discussed previously.

Figure 4 presents, similarly to Fig. 1, the intercomparison of sea and ice surface temperature (called here SST, which includes IST) from the datasets providing this variable: ERA5 at 10:00 am/10:00 pm LT, the IASI-CDR, IASI-NN (at 09:30 am/09:30 pm LT), and the all sky, all day ESA CCI/C3S SST (labeled here simply by ESA CCI/C3S) shown in dotted black line. It is the same for the day and night panels as it is a combined product. The ERA5 land-sea mask is applied to the first three global datasets. The upper panel shows the global mean SST, and the lower panels display day- and night-time anomalies. We note that the y-axis ranges for both the time series and anomalies in Fig. 4 are smaller than those in Fig. 1 for LST, consistent with the lower variability of SST relative to land temperatures. The deseasonalized anomaly time series (Fig. 1c and d) show strong consistency across all datasets in both timing and amplitude of interannual variability, including the 2015–2016 El Niño warming and the subsequent 2020–2022 La Niña cooling. Larger inter-dataset differences, in particular post-2020 for IASI (CDR and NN), reflect systematic offsets attributable to differences in clear-sky sampling strategy rather than spurious trends or variability. The stronger negative anomalies in IASI during 2020–2022 are consistent with the La Niña period, during which increased cloudiness in the tropics reduces the availability of clear-sky retrievals, causing the IASI global mean to be cooler. The ESA CCI/C3S SST product shows systematically damped anomaly peaks in both the positive and negative directions relative to the clear-sky datasets, consistent with the spatial and temporal smoothing inherent to the optimal interpolation scheme used to produce gap-free, all sky, all day coverage, which reduces the amplitude of extreme anomalies by construction.

Biases between the IASI-NN SST product (all sky, all day average) and IASI-CDR, ERA5, and ESA CCI/C3S are shown in Fig. 5 (for IASI, Metop-A and Metop-B are averaged together). We note that the biases between the datasets are much smaller than the land ones (Figs. 1 and 2) and similar between daytime and nighttime, primarily because the ocean surface is more homogeneous. Over the open ocean, IASI-NN agrees closely with all three products: global mean biases are small (+0.01 to +0.31 K), RMSE stays below 1 K in every panel, and differences across most of the ice-free ocean lie within ±1 K. Thin warm filaments along the western boundary currents (Gulf Stream, Kuroshio) mark sharp SST fronts that are smoothed in the gap-filled ESA CCI/C3S analysis. The largest discrepancies are confined to high latitudes and differ in sign between products. Sharing the same instrument and the same clear-sky sampling, IASI-NN and IASI-CDR isolate the effect of the retrieval method: they show the closest agreement of all pairs (RMSE < 0.9 K), with IASI-NN colder than IASI-CDR over the Arctic, consistent with IASI-CDR being trained on, and therefore not independent of, ERA5 (Sect. 2.7). Relative to ERA5, IASI-NN is colder over the Southern Ocean, in line with ERA5's documented warm bias over sea ice (Graham et al., 2019; Herrmannsdörfer et al., 2023); relative to the all-sky ESA CCI/C3S analysis it is instead warmer across the Arctic. These contrasting signs indicate that the high-latitude spread reflects the differing definition, retrieval and sea-ice treatment of skin temperature across the datasets (Sect. 2.7) rather than a single common cause.