Multiscale Evaluation and Error Characterization of HY-2B Fused Sea Surface Temperature Data

Abstract

The Haiyang-2B (HY-2B) satellite, launched on 25 October 2018, carries both active and passive microwave sensors, including a scanning microwave Radiometer (SMR), to deliver high-precision, all-weather global observations. Sea surface temperature (SST) is among its key products. We evaluated the HY-2B SMR Level-4A (L4A) SST (25 km resolution) over the North Pacific (0–60°N, 120°E–100°W) for the period 1 October 2023 to 31 March 2025 using the extended triple collocation (ETC) and dual-pairing methods. These comparisons were made against the Remote Sensing System (RSS) microwave and infrared (MWIR) fused SST product and the National Oceanic and Atmospheric Administration (NOAA) in situ SST Quality Monitor (iQuam) observations. Relative to iQuam, HY-2B SST has a mean bias of –0.002 °C and a root mean square error (RMSE) of 0.279 °C. Compared to the MWIR product, the mean bias is 0.009 °C with an RMSE of 0.270 °C, indicating high accuracy. ETC yields an equivalent standard deviation (ESD) of 0.163 °C for HY-2B, compared to 0.157 °C for iQuam and 0.196 °C for MWIR. Platform-specific ESDs are lowest for drifters (0.124 °C) and tropical moored buoys (0.088 °C) and highest for ship and coastal moored buoys (both 0.238 °C). Both the HY-2B and MWIR products exhibit increasing ESD and RMSE toward higher latitudes, primarily driven by stronger winds, higher columnar water vapor, and elevated cloud liquid water. Overall, HY-2B SST performs reliably under most conditions, but incurs larger errors under extreme environments. This analysis provides a robust basis for its application and future refinement.

1. Introduction

Sea surface temperature (SST) is a critical parameter for monitoring, understanding, and forecasting heat, momentum, and gas fluxes across various spatial and temporal scales [1,2,3,4,5]. These fluxes govern the intricate coupling between the atmosphere and the ocean. SST not only influences air–sea moisture and heat exchange but also serves as a fundamental indicator of ocean circulation patterns, frontal zones, water mass distributions, and other dynamic processes [6,7,8]. SST data are primarily obtained from in situ measurements (ships and buoys) and satellite remote sensing. In situ measurements offer high accuracy but have limited spatial coverage and resolution, and such datasets often exhibit quality and completeness issues due to occasional outliers [9]. In contrast, satellite-derived SST products provide high spatiotemporal resolution and near-synoptic global coverage, making them increasingly important in oceanographic research.

Spaceborne scanning microwave radiometers provide continuous, all-weather observations and have been used to generate a global SST record for over 40 years [10,11]. However, differences in instrument design and calibration among microwave radiometers can introduce biases in brightness temperature measurements, affecting the accuracy of SST retrievals [12,13]. In addition, sensor performance can degrade over time, leading to time-dependent errors in long-term SST observations [14]. Therefore, obtaining more comprehensive and accurate SST products requires fusing observations from multiple satellite sensors, which improves spatial coverage and consistency.

The Haiyang-2B (HY-2B) satellite, launched in October 2018, carries the Scanning Microwave Radiometer (SMR), which continuously measures SST, wind speed, atmospheric water vapor, cloud liquid water, and rainfall intensity [15,16]. The addition of SMR significantly expands the spatiotemporal coverage and timeliness of SST observations. Previous studies have validated the accuracy of HY-2B Level-2B (L2B) SST products. Zhou et al. [17] used NOAA iQuam in situ data to assess HY-2B SST, reporting a bias of –0.02 °C and an RMSE of 0.80 °C. Similarly, Zhang et al. [18] reported a bias of 0.09 °C and an RMSE of 0.72 °C for these products.

Since its launch, HY-2B has been operational for over six years, far exceeding its design life. However, no validation study has yet focused on the HY-2B L4A SST fusion product. Compared to traditional single-sensor products, fused SST products offer substantially improved spatial coverage, enabling near-global observations. In generating fused SST products, different institutions may combine varied data sources (e.g., infrared SST, microwave SST, and in situ data) and processing methods to meet specific objectives. These sources differ in format, spatial/temporal resolution, cloud detection algorithms, and quality control procedures. Additionally, differences in fusion algorithms, initial field generation, land–sea boundary handling, and ice-masking techniques can lead to discrepancies in the final product performance [19].

In this study, we apply extended triple collocation (ETC) and direct pairwise comparison methods to evaluate the operational HY-2B L4A fused SST product for the period from 1 October 2023 to 31 March 2025. Section 2 describes the data and methods. Section 3 and Section 4 present the results and discussion, respectively. Section 5 concludes the study. The flowchart of the step-by-step research is shown in Figure 1.

2. Materials and Methods

This study employs multiple remote sensing SST datasets along with comparative analyses and multi-scale validation to assess the observational performance of the HY-2B SMR over the North Pacific. The datasets include the HY-2B L4A fused SST product, the RSS MWIR fused SST product, NOAA iQuam in situ observations for accuracy assessment, and ERA5 reanalysis data for environmental context analysis.

2.1. Study Area

The study area covers the North Pacific (0–60°N, 120°E–100°W) (Figure 2). This region spans from the tropics to mid and high latitudes and includes dynamic features such as the Equatorial Warm Pool, the Kuroshio and California Currents, and the North Pacific Subtropical High. Sea surface temperature in the North Pacific exhibits pronounced spatio-temporal variability and strong seasonality, making it an ideal region for evaluating satellite SST accuracy and stability across diverse climate regimes [20,21]. Moreover, a dense network of buoys in this area provides ample collocated observations for validation.

2.2. Datasets

2.2.1. HY-2B SST

The HY-2B L4A fused sea surface temperature (SST) product used in this study is an operational daily (24 h) fusion field produced by the National Satellite Ocean Application Service (NSOAS) and distributed via the NSOAS Ocean Dynamics portal. The exact product employed here is “MUL_OPER_SST_L4A_FU_01D_20250101_dps_250_10_sst” (product level L4A). Files are provided in NetCDF-4 format with a nominal spatial resolution of ~25 km; the file name convention indicates a daily aggregation (“01D”), which we used for all collocations and analyses. The L4A fusion combines multiple satellite retrievals (microwave and infrared) through NSOAS’s operational fusion procedures (spatiotemporal interpolation, quality control, bias correction and weighting) to produce continuous, gap-filled SST fields suitable for large-scale validation and applications. We regridded/processed the daily HY-2B L4A fields as described in Section 2.3. to match the temporal and spatial collocation scheme used for MWIR and iQuam. The HY-2B SST dataset spans October 2023–March 2025.

2.2.2. MWIR SST

The MWIR product (version 5.1) is a daily fused SST dataset provided by Remote Sensing Systems (RSSs). It employs an optimized interpolation scheme to map daily SST estimates onto a 0.09° global grid, with records covering 2002 to the present. MWIR integrates data from multiple sensors: microwave radiometers, including the GPM Microwave Imager (GMI), TRMM Microwave Imager (TMI), Advanced Microwave Scanning Radiometer for EOS (AMSR-E), Advanced Microwave Scanning Radiometer 2 (AMSR2), and WindSat, and infrared (IR) radiometers, including Moderate Resolution Imaging Spectroradiometer (MODIS), on the Aqua and Terra satellites and Visible Infrared Imaging Radiometer Suite (VIIRS) on Suomi NPP and NOAA-20. IR sensors provide high spatial resolution but require clear-sky conditions, whereas microwave sensors can retrieve SST through clouds. In that the MWIR dataset does not make use of HY-2B, it serves as an independent reference. A diurnal cycle correction is applied using modeled base temperature adjustments to mitigate daily thermal expansion/contraction effects. Although MWIR does not directly ingest in situ data, its constituent microwave sensors are buoy-calibrated. The MWIR dataset serves as an independent reference for cross-comparing the HY-2B fusion SST product. MWIR data are available from the RSS data repository.

2.2.3. In Situ Observation Data

High-precision in situ SST observations are essential for satellite SST validation. The NOAA in situ SST Quality Monitor (iQuam) dataset, developed by NOAA’s Satellite Applications and Research Center, is widely used for this purpose. The current iQuam version (2.10) provides quality control (QC) of in situ SST measurements, online monitoring of QC-processed data, and reformatted SST output with quality flags [22]. iQuam incorporates SST data from commercial vessels, drifting buoys, tropical moored buoys (T-M), coastal moored buoys (C-M), and Argo floats, and is extensively used for satellite validation [23,24,25].

The iQuam QC procedures include pre-screening, plausibility checks, internal consistency checks, and cross-consistency checks. Each observation is assigned a quality level from 1 (poorest) to 5 (highest); this study uses only level 5 (highest quality) observations from ships, drifting buoys, and moored buoys. All iQuam data are provided in NetCDF format. Xu et al. used triple collocation to estimate platform-specific observation errors in iQuam [26]: their results (standard deviations of 0.75 K for ships, 0.21–0.22 K for drifting buoys/Argo, 0.17 K for tropical moorings, and 0.40 K for coastal moorings) indicate relatively low random errors across all platforms. This underscores the reliability of high-quality iQuam data for validation.

2.2.4. ERA5 Reanalysis Data

ERA5 is the fifth-generation global climate reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF) under the Copernicus Climate Change Service (C3S), superseding ERA-Interim [27,28]. ERA5 provides hourly global fields on a ~31 km horizontal grid with 137 vertical levels, encompassing over 30 atmospheric, land, and oceanic variables. In this study, we employ daily-mean ERA5 single-layer data for auxiliary environmental parameters: total column water vapor, total column cloud liquid water, and the 10 m eastward (U) and northward (V) wind components. Notably, ERA5 assimilates a wide range of observations (over 200 data types, including satellite and conventional upper-air measurements) but excludes HY-2B sensor data. It also directly assimilates in situ ship and buoy observations (e.g., 10 m wind, 2 m humidity, sea surface pressure) [28]. These ERA5 variables are used to analyze HY-2B SST error characteristics under varying environmental conditions.

2.3. Methods

2.3.1. Data Matching and Quality Control

To ensure comparability between the HY-2B, MWIR, and iQuam datasets, we employ a unified spatiotemporal collocation and quality control framework. First, the MWIR SST dataset was re-gridded to a 0.25° latitude–longitude grid using two-dimensional spatial interpolation. For spatial collocation, each satellite grid point must lie within 25 km of an in situ observation. Hourly in situ observations are averaged to daily means to match the temporal resolution of the satellite products. Within each 0.25° grid cell and day, all available observations from ships, drifting buoys, tropical moored (T-M) buoys, and coastal moored (C-M) buoys are collocated.

All initial matchups undergo strict quality control: any records with missing values or SST outside the 0–34 °C range are discarded, and only iQuam observations with quality level 5 (highest quality) are retained. Figure 3 illustrates the spatial distribution of the resulting collocated points. Finally, we tally the number of samples by platform type to form platform-specific datasets for subsequent analysis. These collocation and QC procedures ensure consistency and reliability across all SST data sources.

2.3.2. Comparison Methods

Comparing satellite-retrieved SST with in situ SST from ships and buoys is a standard approach for evaluating retrieval accuracy [29,30,31,32]. Such comparisons assume that in situ observations represent the true sea surface temperature; however, in situ measurements contain both random and systematic errors [33,34]. Moreover, satellite sensors observe different layers of the surface: infrared radiometers sense the near-surface “skin” temperature (top few micrometers), while microwave radiometers retrieve the sub-skin SST (roughly 1 mm depth) [35,36,37]. Cold-skin and diurnal warming effects can create significant temperature differences between the skin, sub-skin, and bulk layers [38,39]. These factors introduce inherent biases and uncertainties in direct comparisons.

To address these limitations, triple collocation (TC) methods have been developed to estimate the error variances of three mutually independent datasets without assuming any single dataset as truth [40,41]. TC has been widely applied to validate soil moisture, precipitation, sea surface salinity, sea surface temperature, and sea surface wind and waves [42,43,44,45,46,47,48]. Building on the TC method, the extended triple collocation (ETC) method was proposed. ETC adopts the same basic assumptions as TC but derives additional performance metrics, including the correlation coefficient between each product and the unknown target, as well as the scaled unbiased signal-to-noise ratio (${\mathrm{SNR}}_{\mathrm{sub}}$). Compared to standard deviation, ETC provides a complementary perspective on product performance [49].

2.3.3. Statistical Indicators and Analysis

To quantify SST accuracy, we compute the mean bias, root mean square error (RMSE), and Pearson correlation coefficient (R) between each satellite product and the reference (in situ) SST. Bias is defined as the average difference between the satellite SST and the reference SST. RMSE represents the standard deviation of these differences and is sensitive to outliers, serving as a robust measure of overall precision. The Pearson correlation coefficient R indicates the strength of the linear relationship between the satellite and reference SST values. The mathematical formulas for these metrics are given by

$$\mathrm{Bias}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N(S_i-O_i)}$$

(1)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{N}}{{\displaystyle \sum _{i=1}^N(S_i-O_i)}}^2}$$

(2)

$$R={\displaystyle \frac{{\displaystyle \sum _{i=1}^N(S_i-\overline{S})}(O_i-\overline{O})}{\sqrt{{{\displaystyle \sum _{i=1}^N(S_i-\overline{S})}}^2 \times {\displaystyle \sum _{i=1}^N(}{O}_i-\overline{O}{)}^2}}}$$

(3)

In the formula, N denotes the total number of matching data points between remote sensing observations of SST and iQuam in situ SST, while S and O denote the satellite data set and reference data set being verified, respectively.

2.3.4. Extended Triple Collocation (ETC)

ETC is a technique for estimating the noise error variance (errVar) and correlation coefficient (rho) of three measurement systems (e.g., satellite, in situ, and model-based products) relative to the unknown true values of the measured variables (e.g., sea surface temperature, soil moisture, wind speed) [49]. ETC applies to both absolute and anomaly values without requiring rescaling to a reference system.

Traditional approaches (e.g., linear regression) assume in situ data as truth, yet these observations contain both random and systematic errors. Therefore, using extended triple collocation (ETC) allows for independent estimation of the error variance for each measurement system. ETC enables independent estimation of each system’s error variance, avoiding reliance on a single reference dataset. Unlike traditional methods, ETC does not assume any dataset as truth, thus providing independent error variance estimates for all systems [50]. ETC uses the same assumptions as TC but provides an additional validation parameter, namely the coefficient of determination (${\rho }_{{\mathrm{t}, \mathrm{X}}_i}^2$) relative to the unknown “true” value of the measured variable.

Assume that three independent measurement values are linearly related to the unknown true value T. The affine error model between the measurement values and T can be expressed as follows:

$$X_i=X_i^{\prime }+{\epsilon }_i={\alpha }_i+{\beta }_{i}t+{\epsilon }_i$$

(4)

where $X_i(i\in \{1,2,3\})$ is a set of three spatially and temporally co-located datasets (${\mathrm{X}}_1$: system-based products; ${\mathrm{X}}_2$: model-based products; ${\mathrm{X}}_3$: in situ products), each with additive random error ${\epsilon }_i$; $t$ is the unknown true value (or true state) and ${\alpha }_i$ and ${\beta }_i$ are the multiplicative bias and proportionality factor of dataset $X_i$ relative to the true value.

In this study, we adopted the covariance combination principle among the three products without rescaling [40,49]. The covariance between different products is

$$\mathrm{Cov}(X_i, X_j)=\mathrm{E}(X_i{X}_j)-\mathrm{E}(X_i)\mathrm{E}(X_j)={\beta }_i{\beta }_j{\sigma }_t^2+{\beta }_i\mathrm{Cov}(t,{\epsilon }_j)+{\beta }_j\mathrm{Cov}(t,{\epsilon }_i)+\mathrm{Cov}({\epsilon }_i,{\epsilon }_j)$$

(5)

$${\sigma }_t^2=\mathrm{Var}(\mathrm{t})$$

(6)

The basic assumptions of ETC are that (i) the expected value of errors from independent data sources is zero $(\mathrm{E}({\epsilon }_i)=0)$, (ii) they are mutually independent $(\mathrm{Cov}({\epsilon }_i,{\epsilon }_j)=0, \mathrm{for} \mathrm{i}\ne \mathrm{j})$, and (iii) the errors of the three products are independent of each other and unrelated to the true value $t$ $(\mathrm{Cov}(t,{\epsilon }_i)=0)$.

The covariance of these three products can be simplified as follows:

$${\mathrm{Q}}_{ij}\equiv \mathrm{Cov}(X_i, X_j)=\left\{\begin{array}{ll}{\beta }_i {\beta }_j {\sigma }_t^2,\hfill & \mathrm{for} i\ne j\hfill \\ {\beta }_i^2 {\sigma }_t^2+{\sigma }_{{\epsilon }_i}^2,\hfill & \mathrm{for} i=j\hfill \end{array}\right.$$

(7)

The error variance ${{\sigma }_{{\epsilon }_i}}^2=\mathrm{Var}({\epsilon }_i)$. The error standard deviation (ESD) of the three products can be determined from the square root of the error variance

$${\sigma }_{{\epsilon }_1}=\left[\begin{array}{c}\sqrt{ Q_{11}-{\displaystyle \frac{Q_{12} Q_{13}}{Q_{23}}} }\\ \sqrt{ Q_{22}-{\displaystyle \frac{Q_{12} Q_{23}}{Q_{13}}} }\\ \sqrt{ Q_{33}-{\displaystyle \frac{Q_{13} Q_{23}}{Q_{12}}} }\end{array}\right]$$

(8)

The correlation coefficient between t and X expressed in terms of covariance values is

$${\rho }_{t,X}=\pm \left[\begin{array}{c}\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{12} {\mathrm{Q}}_{13}}{ {\mathrm{Q}}_{11} {\mathrm{Q}}_{23} }}}\\ {\mathrm{sign}(\mathrm{Q}}_{13} {\mathrm{Q}}_{23})\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{12} {\mathrm{Q}}_{23}}{ {\mathrm{Q}}_{22} {\mathrm{Q}}_{13} }}}\\ {\mathrm{sign}(\mathrm{Q}}_{12} {\mathrm{Q}}_{23})\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{13} {\mathrm{Q}}_{23}}{ {\mathrm{Q}}_{33} {\mathrm{Q}}_{12} }}}\end{array}\right]$$

(9)

It is worth noting that the ${\rho }_{t,X}$ provided by ETC has signed ambiguity; however, in practical applications, ${\rho }_{t,X}$ is always positive. To describe the combined effects of product sensitivity $({\beta }_i)$, true signal variability $({\sigma }_t)$, and measurement error variability $({\sigma }_{\epsilon })$, the formula for calculating the correlation coefficient squared (the scaled unbiased signal-to-noise ratio ${\mathrm{SNR}}_{\mathrm{sub}}$) is as follows:

$${\mathrm{SNR}}_{\mathrm{sub}}=\pm {\left[\begin{array}{c}\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{12} {\mathrm{Q}}_{13}}{ {\mathrm{Q}}_{11} {\mathrm{Q}}_{23} }}}\\ {\mathrm{sign}(\mathrm{Q}}_{13} {\mathrm{Q}}_{23})\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{12} {\mathrm{Q}}_{23}}{ {\mathrm{Q}}_{22} {\mathrm{Q}}_{13} }}}\\ {\mathrm{sign}(\mathrm{Q}}_{12} {\mathrm{Q}}_{23})\sqrt{{\displaystyle \frac{{\mathrm{Q}}_{13} {\mathrm{Q}}_{23}}{ {\mathrm{Q}}_{33} {\mathrm{Q}}_{12} }}}\end{array}\right]}^2$$

(10)

where ${\mathrm{SNR}}_{\mathrm{sub}}$ represents the proportional unbiased signal-to-noise ratio. It contains information about product sensitivity and can be used to evaluate whether the noise level of the system is suitable for detecting changes in target variables. We apply ETC to three datasets—HY-2B, MWIR, and iQuam—across four platforms (Ship, Drifter, T-M, and C-M) to derive their random errors (ESD) and ${\mathrm{SNR}}_{\mathrm{sub}}$ values. To elucidate error sources, we examine the interplay between sensor characteristics and environmental conditions: infrared (IR) sensors offer high spatial resolution but are hindered by cloud cover and atmospheric water vapor, whereas microwave sensors provide all-weather SST retrievals at the cost of coarser resolution and greater susceptibility to surface roughness and wind waves. The MWIR product integrates data from two thermal-infrared sensors (MODIS, AVHRR) and three passive microwave sensors (AMSR-E, AMSR2, WindSat) to optimize the trade-off between observational accuracy and spatial coverage [51].

3. Results

3.1. Comparison with In Situ SST

Figure 4 shows scatter plots and bias histograms comparing the HY-2B and MWIR SST products with the iQuam in situ observations over the period of October 2023–March 2025. In each scatter plot, the data cluster tightly around the 1:1 line, with regression slopes very close to unity and coefficients of determination (R²) exceeding 0.99. The bias histograms for both products are approximately normal and centered near zero, with most differences within ±0.5 °C. These results demonstrate strong consistency between the satellite SST products and the in situ measurements.

Table 1. Bias, RMSE, and R² of HY-2B and RSS MWIR SST against iQuam in situ observations by platform.

Match Type	Indicator	iQuam Platforms
		All	Ship	Drifter	T-M	C-M
HY-2B VS iQuam	Bias (°C)	−0.002	−0.090	0.042	0.007	−0.063
	RMSE (°C)	0.279	0.427	0.177	0.145	0.376
	R-squared	0.9987	0.9974	0.9994	0.9903	0.9969
	Number	5,686,212	535,382	3,359,875	126,826	1,663,980
MWIR VS iQuam	Bias (°C)	−0.011	−0.017	0.002	−0.057	−0.042
	RMSE°C	0.253	0.337	0.205	0.175	0.350
	R-squared	0.9989	0.9983	0.9991	0.9871	0.9970
	Number	5,252,568	200,550	3,635,475	135,713	1,280,670

summarizes the comparison statistics (bias, RMSE, R²) for HY-2B and MWIR versus iQuam, broken down by platform (All, Ship, Drifter, T-M, C-M). For HY-2B versus iQuam (all platforms), the mean bias is –0.002 °C, RMSE is 0.279 °C, and R² is 0.9987. The lowest RMSE values are found for drifters (0.177 °C) and tropical moored buoys (0.145 °C), with corresponding R² values of 0.9994 and 0.9903. In contrast, ship and coastal moored observations yield higher RMSEs (0.427 °C and 0.376 °C, respectively). For MWIR versus iQuam, the overall bias is –0.011 °C and RMSE is 0.253 °C. MWIR also performs best on drifters (RMSE 0.205 °C) and tropical moored buoys (0.175 °C), while ships and coastal moorings have RMSEs of 0.337 °C and 0.350 °C, respectively. Both satellite products achieve similar overall accuracy (R² > 0.99), performing particularly well on drifter and tropical buoy platforms.

These results suggest that drifter and tropical moored buoys, which sample relatively uniform open-ocean waters, yield the highest agreement with satellite SST (lowest representativeness error). By contrast, ships and coastal moorings often operate in regions with strong SST gradients or local disturbances, leading to larger point errors. In those heterogeneous environments, representativeness errors increase. Notably, on drifter and T m platforms, HY-2B exhibits the smallest random errors, whereas its errors are larger for ship-based data. Overall, the HY-2B and MWIR products demonstrate comparable accuracy (R² > 0.99), but platform-specific differences highlight distinct observational characteristics.

Figure 5 displays the spatial distribution of HY-2B SST error metrics (bias, RMSE, and Pearson’s R) on a 2° × 2° grid (HY-2B vs. iQuam). In mid- to high-latitude open-ocean regions, biases cluster near zero and RMSE remains below 0.25 °C, with R exceeding 0.99, indicating excellent agreement where sea conditions are homogeneous and sampling is dense. In contrast, coastal zones and data-sparse equatorial areas exhibit pronounced negative biases (down to –0.1 °C), elevated RMSE (up to 0.4 °C), and slightly reduced R (around 0.95), reflecting increased representativeness error and local disturbances. Notably, these spatial patterns show no clear seasonal drift: overall, bias and RMSE stay low and uniform in well-sampled open waters, whereas accuracy degrades in heterogeneous or sparsely observed regions.

3.2. Comparison of Fusion SST Products

Figure 6 compares the HY-2B and RSS MWIR SST fusion products directly. The scatter density plot shows the HY-2B SST on the x-axis versus the MWIR SST on the y-axis. The data points cluster tightly around the 1:1 reference line. A linear fit yields y = 0.99x + 0.15 °C with R² = 0.999, indicating extremely strong agreement. The overall mean bias (HY-2B minus MWIR) is only 0.009 °C, with an RMSE of 0.287 °C. The bias distribution is approximately normal and centered near zero, with the majority of errors within ±0.5 °C. In summary, HY-2B and MWIR fusion products show negligible systematic and random differences under large-sample conditions, confirming their exceptional consistency and reliability.

Figure 7 and Figure 8 present monthly maps of the HY-2B minus MWIR bias from October 2023 to March 2025. In general, these maps reveal no systematic seasonal drift. Equatorial regions remain near zero bias year-round, with no obvious zonal trends. Small alternating positive/negative bias patterns appear in mid- and high-latitude seas. For example, the Sea of Okhotsk exhibits a positive bias from January to July and a negative bias from August to December. During boreal summer (June–August), bias hotspots emerge in the northwestern Pacific, the Sea of Okhotsk, and the Sea of Japan—likely associated with the Kuroshio Extension and other regional dynamics [52]. In boreal winter (December–February), these regions show negative biases (HY-2B cooler than MWIR). Overall, biases are concentrated in high-latitude and coastal areas, while open-ocean biases remain minimal and uniform. These monthly maps confirm the absence of any significant seasonal drift: errors appear randomly distributed, and the relatively short time span of this study precludes definitive identification of longer-term seasonal or regional biases.

3.3. ETC Analysis Results

Extended triple collocation (ETC) was used to evaluate the performance of the three independent SST datasets (HY-2B, MWIR, and iQuam) on each platform. ETC separates the random error variance of each system from representativeness error. The resulting error standard deviation (ESD) quantifies the unbiased random error, and the sub-sampled signal-to-noise ratio (${\mathrm{SNR}}_{\mathrm{sub}}$) indicates the relative strength of the true signal.

Table 2. ETC analysis results of HY-2B, MWIR, and iQuam on different platforms.

Data	ETC	ALL	Ship	Drifting	T-M	C-M
HY-2B	ESD (°C)	0.163	0.238	0.120	0.088	0.238
	${\mathrm{SNR}}_{\mathrm{sub}}$	0.9996	0.9991	0.9997	0.9962	0.9986
MWIR	ESD (°C)	0.196	0.252	0.164	0.115	0.267
	${\mathrm{SNR}}_{\mathrm{sub}}$	0.9993	0.9990	0.9994	0.9933	0.9983
iQuam	ESD (°C)	0.157	0.220	0.123	0.111	0.218
	${\mathrm{SNR}}_{\mathrm{sub}}$	0.9996	0.9992	0.9997	0.9941	0.9988

lists the ETC analysis results for HY-2B, MWIR, and iQuam data across various platforms. From this, we observe a significant difference between ETC-based ESD and conventional RMSE, as ESD excludes the influence of representativeness errors [53,54]. At the aggregate level (all platforms), iQuam has the smallest ESD (0.157 °C), followed by HY-2B (0.163 °C), and MWIR has the largest (0.196 °C). The corresponding ${\mathrm{SNR}}_{\mathrm{sub}}$ values are very close to 1 (0.9996 for iQuam and HY-2B, 0.9993 for MWIR), indicating that all three systems have strong detection capability for true SST variations.

Platform-specific results reveal that the lowest errors occur on drifter and tropical moored buoy platforms. For drifters and tropical moorings (T-M), the ESDs are as follows: iQuam = 0.123 °C and 0.111 °C (T-M); HY-2B = 0.120 °C and 0.088 °C; MWIR = 0.164 °C and 0.115 °C, respectively. In contrast, ships and coastal moorings show higher ESD values for all datasets. This pattern likely arises because drifters and T-M buoys are generally located in the open ocean, where conditions are more stable and representativeness errors are low. Notably, on the drifter and T-M platforms, HY-2B’s ESD is slightly lower than iQuam’s, indicating that HY-2B SST achieves exceptionally high precision under those conditions. Overall, iQuam exhibits the highest accuracy (smallest ESD) as expected, while HY-2B provides better accuracy than MWIR across all platforms.

4. Discussion

Our results indicate that the HY-2B L4A fusion SST product performs exceptionally well. The scatter plots (Figure 4 and Figure 6) show that both HY-2B and MWIR products are highly consistent with each other and with in situ observations, with most errors randomly distributed within ±0.5 °C. In situ validation reveals that HY-2B has essentially zero mean bias over the study period and a low RMSE (~0.2–0.3 °C). This represents a substantial improvement over earlier HY-2B Level-2B products: for example, previous work reported a bias of 0.1 °C and an RMSE of 0.87 °C for the HY-2B L2 SST [17]. Thus, the L4 fusion processing appears to significantly reduce systematic errors and improve accuracy. Overall, the HY-2B L4A product exhibits very small systematic errors and high correlation with in situ SST.

Previous studies have shown that environmental parameters such as sea surface wind speed, atmospheric water vapor, and cloud liquid water can introduce errors in SST retrieval. As sea surface wind speed increases, sea surface roughness and foam increase, altering microwave emission characteristics and thereby increasing SST retrieval uncertainty. Additionally, water vapor and cloud liquid water in the atmosphere both absorb and scatter microwave signals, increasing the complexity of atmospheric correction. These factors can induce inter-channel crosstalk effects, significantly impacting SST retrieval accuracy [55].

We analyzed the variation patterns of HY-2B SST error metrics (ESD, ${\mathrm{SNR}}_{\mathrm{sub}}$, bias, and RMSE) with respect to environmental variables (time, latitude, sea surface temperature, wind speed, columnar water vapor, and columnar cloud liquid water) to comprehensively evaluate its performance.

4.1. Time-Dependent Characteristics of SST Errors

Figure 9 shows the evolution of error metrics from October 2023 to March 2025 (10-day averages). All products remain generally stable over time. In Figure 9a, HY-2B and iQuam maintain low ESD (<0.2 °C) and high ${\mathrm{SNR}}_{\mathrm{sub}}$ (>0.999) throughout the period, with their curves nearly overlapping. MWIR has a slightly higher ESD and marginally lower ${\mathrm{SNR}}_{\mathrm{sub}}$, but its temporal trend closely follows HY-2B. Figure 9b shows that ${\mathrm{SNR}}_{\mathrm{sub}}$ stays above 0.998 for all products, indicating a consistently high degree of agreement.

Figure 9c,d indicate that the HY-2B bias relative to iQuam fluctuates around 0 °C with no obvious trend, and RMSE fluctuates around 0.2 °C. The bias and RMSE time series for MWIR are similarly flat. Notably, both satellite products exhibit somewhat higher ESD and RMSE during boreal summer (June–August); MWIR in particular shows a noticeable bias increase in that period, possibly due to stronger solar heating affecting the sensors. Over the entire period, however, the consistency among HY-2B, MWIR, and iQuam remains high. This suggests that both satellite products have stable time-series quality, with no significant drift or degradation over the six-month span.

4.2. SST Error Characteristics at Different Latitudes

Figure 10 shows error metrics as a function of latitude (0–58°N). In Figure 10a, the ESD for all three products increases with latitude from 0° up to about 35°N. Beyond 35°N, the MWIR ESD remains roughly between 0.2 and 0.3 °C, while iQuam and HY-2B ESD gently decline with further latitude. Figure 10b shows that ${\mathrm{SNR}}_{\mathrm{sub}}$ stays above 0.99 at all latitudes, with only small dips around 8°N and 44°N. Above 50°N, ${\mathrm{SNR}}_{\mathrm{sub}}$ fluctuates more significantly for all products, reflecting increased random error. These high-latitude fluctuations are likely due to stronger winds and more variable conditions, which degrade SST accuracy.

Figure 10c indicates that HY-2B bias (relative to iQuam) closely mirrors MWIR bias across latitudes: slight positive biases occur in low latitudes and slight negative biases in mid-to-high latitudes. Over 0–55°N, the difference in bias between HY-2B and MWIR remains within ±0.1 °C. Figure 10d shows that both satellite products have increasing RMSE (and ESD) with latitude up to 35°N, then a gradual decrease. The similarity in latitudinal error trends suggests that both products respond similarly to environmental changes with latitude. In summary, SST errors generally grow with latitude into mid-latitudes due to harsher conditions (higher winds, etc.) but then stabilize, and both HY-2B and MWIR exhibit comparable error behavior as a function of latitude.

4.3. Changes in SST Error Characteristics at Different Temperatures

Figure 11 shows the changes in error characteristics within the 2 °C in situ sea surface temperature interval, ranging from 0 to 30 °C. Since there are few data matching points above 30 °C, these data will be excluded in this paper. The ESD changes of HY-2B and iQuam are similar. Within the 0 to 14 °C range, the ESD of both systems increases with rising temperature. In contrast, MWIR starts at a higher ESD value (0.3 °C). Subsequently, the ESD of all three products decreases with increasing sea surface temperature, from about 0.25 °C to 0.1 °C. Where HY-2B and iQuam have similar ESD values, both lower than MWIR, it indicates that the random error of HY-2B is smaller than that of MWIR in the low temperature range. The ${\mathrm{SNR}}_{\mathrm{sub}}$ of iQuam and HY-2B changes similarly with SST. Within the SST range of 0 to 14 °C, ${\mathrm{SNR}}_{\mathrm{sub}}$ decreases with increasing SST, while MWIR’s ${\mathrm{SNR}}_{\mathrm{sub}}$ slightly increases with rising SST. However, as sea surface temperature increases, the ${\mathrm{SNR}}_{\mathrm{sub}}$ of all three products gradually increases to above 0.99. As shown in Figure 11, under extreme high-temperature conditions, the errors of all three products increase.

As shown in Figure 11c,d, within the range excluding extreme low and high temperatures, the bias of HY-2B and MWIR relative to iQuam SST remains around 0 degrees. Compared to HY-2B, MWIR has a higher RMSE relative to iQuam at low temperatures (0.35 °C), but as SST increases, the two become increasingly similar. The overall performance is similar to the ESD variation, with both showing a decreasing trend with increasing SST after 14 °C.

4.4. Variations in SST Error Under Wind Speed Influence

Sea surface wind speed is an important environmental parameter affecting the accuracy of sea surface temperature measurements, largely due to the “cold skin effect” [36]. High wind speeds increase water mixing, thereby reducing temperature differences between different depths. Additionally, high sea surface wind speeds can cause white caps, thereby increasing the uncertainty of sea temperature measurements.

Figure 12 shows error metrics as a function of surface wind speed (0–19 m/s). As wind increases, MWIR’s ESD begins to rise noticeably above 10 m/s, and its ${\mathrm{SNR}}_{\mathrm{sub}}$ correspondingly decreases (Figure 12a,b). By contrast, HY-2B and iQuam show no obvious trend in ESD or ${\mathrm{SNR}}_{\mathrm{sub}}$ with wind speed; both remain relatively stable with only small fluctuations. Both products’ biases stay near 0 °C across wind speeds (Figure 12c). The RMSE of HY-2B remains low and stable up to 10 m/s, with only slight increases in variability beyond that (Figure 12d). MWIR’s RMSE, however, clearly increases with wind speeds above 10 m/s. These observations are consistent with the cool-skin interpretation [56,57]: increasing wind speeds enhance mixing and reduce the skin–subskin temperature difference, thereby diminishing the cool-skin contribution to satellite minus in situ differences. Since HY-2B SMR retrieves a sub-skin temperature (∼1 mm), the direct cool-skin contribution to HY-2B minus typical in situ near-surface differences is small in our aggregated comparisons; nevertheless, wind-driven roughness and foam remain important contributors to the increased uncertainty at high winds.

4.5. Variations in SST Error Under Columnar Water Vapor Conditions

The attenuation and radiation of water vapor in the atmosphere can affect the signals received by microwave radiometers, thereby increasing the uncertainty in sea surface temperature retrieval [58]. This study investigates the relationship between the SST error characteristics of HY-2B and columnar water vapor (Figure 13). Since columnar water vapor does not directly affect in situ data, the analysis in this section does not include iQuam SST.

Figure 13 analyzes HY-2B and MWIR errors versus total column water vapor (1–75 mm). Between 1 and 35 mm of vapor, both HY-2B and MWIR ESD decrease, reaching about 0.14 °C (Figure 13a), after which the ESD remains roughly constant; minor oscillations occur at very high vapor (>60 mm). The ${\mathrm{SNR}}_{\mathrm{sub}}$ for both products (Figure 13b) drops when vapor is below 10 mm or above 40 mm (reaching minima) and then increases again at very high vapor (>65 mm). In terms of bias, HY-2B tends to have a slight positive bias in the mid range (11–55 mm), whereas MWIR shows a slight negative bias at very low and very high vapor, being near zero otherwise (Figure 13c). The RMSE trends (Figure 13d) largely mirror those of ESD: RMSE declines with increasing vapor up to 35 mm and then fluctuates. These patterns suggest that moderate amounts of water vapor improve retrieval (reducing random error), but extremes of low or high vapor content degrade accuracy through increased attenuation.

4.6. Impact of Cloud Liquid Water Content on SST Error Characteristics

Similar to water vapor, the presence of liquid water in clouds also increases the uncertainty of microwave SST retrieval. Figure 14 shows the effect of columnar cloud liquid water (0–1.0 mm) on HY-2B and MWIR SST errors. As liquid water content rises from 0 to ~0.5 mm, the ESD of both HY-2B and MWIR increases substantially (Figure 14a), with MWIR’s ESD exceeding that of HY-2B. Beyond 0.5 mm, the errors tend to saturate or even decrease slightly. Both products maintain generally negative biases as liquid water increases (Figure 14c), and the RMSE (Figure 14d) follows a similar upward trend. This behavior indicates that cloud liquid water exacerbates SST retrieval errors (due to increased microwave attenuation and scattering), although beyond a certain point the incremental effect levels off.

Our analysis indicates that HY-2B SST errors are most pronounced under extreme conditions. For example, at high latitudes (around 60°N), we observe HY-2B biases up to 0.13 °C and RMSE 0.36 °C. The harsher environments there (strong winds, sea ice) contribute to higher retrieval errors. In contrast, tropical and subtropical waters have more uniform conditions (weaker winds, less vapor variability), leading to lower and more stable errors. Overall, HY-2B SST performs reliably under most conditions, but errors increase under conditions of high winds, high latitudes, and heavy atmospheric moisture.

5. Conclusions

This study systematically assessed the observational performance of the HY-2B L4A fusion SST product by comparing it with RSS MWIR satellite SST and NOAA iQuam in situ data from October 2023 to March 2025. All datasets were collocated in time and space for a fair comparison. The HY-2B and iQuam comparison yielded a near-zero mean bias (–0.002 °C), an RMSE of 0.279 °C, and an R² = 0.9987, indicating negligible systematic error and very high accuracy. Platform-specific analysis showed that HY-2B has the lowest RMSE (0.145–0.177 °C) on drifters and tropical moored buoys, outperforming MWIR on those platforms. Ships and coastal moorings exhibited higher RMSE (0.376–0.427 °C) for HY-2B, as expected for these heterogeneous environments. These differences arise because drifters and tropical moorings are located in open ocean with uniform SST, resulting in low representativeness error, whereas ships and coastal moorings often operate in areas with large temperature gradients or local disturbances.

In a direct comparison between the two satellite fusion products, HY-2B and MWIR agree very closely: the average bias is 0.009 °C and the RMSE is 0.270 °C, with most errors within ±0.5 °C. This demonstrates the consistency and reliability of both products under broad conditions. Extended triple collocation analysis confirms these findings. Drifter and tropical moored platforms have the lowest ESD values (e.g., HY-2B ESD of 0.088–0.120 °C), while ships and coastal moorings have higher ESD. Overall, HY-2B’s aggregate ESD (0.163 °C) lies between iQuam (0.157 °C) and MWIR (0.196 °C). The corresponding signal-to-noise ratios are nearly identical for all three products, indicating robust detection of true SST signals. These results suggest that HY-2B has slightly better random error control than MWIR, which contains more noise.

Finally, our analysis revealed that HY-2B SST errors increase under extreme environmental conditions: high winds, high latitudes, and high water vapor or cloud content. For instance, above 50°N, random errors and ${\mathrm{SNR}}_{\mathrm{sub}}$ fluctuations increase due to colder sea and rougher conditions. At wind speeds above 10 m/s, HY-2B RMSE begins to rise (the “cold skin effect”), whereas MWIR errors increase more steadily. Very high or very low total water vapor causes a modest bias increase (0.05 °C) and higher RMSE, reflecting increased attenuation. Similarly, columnar cloud water up to 0.5 mm raises HY-2B and MWIR ESD and RMSE. These findings suggest that atmospheric attenuation and cloud radiation notably affect microwave SST retrieval. Future work could incorporate environmental corrections or advanced multi-source fusion to improve SST accuracy under such challenging conditions.

Overall, the HY-2B L4A fusion SST product demonstrates high accuracy and consistency across most conditions, providing a reliable resource for oceanographic applications.