Data Quality Evaluation for
Eight Global SST Products
Under Typhoon Conditions
Introduction
Under extreme typhoon conditions, localized physical processes and data-blending constraints introduce significant uncertainties into global SST products. This part provides a comprehensive evaluation of eight distinct SST products using statistical metrics (Bias, RMSE, CC) and explores the underlying causes of systemic warm/cold deviations. Furthermore, we dissect how multi-sensor integration algorithms—specifically OISST and OSTIA—manifest representation errors.
1. Statistical Data Quality Evaluation for 8 SST Products
Fig. 1. Metrics of SST. (a) Bias, (b) root-mean-squared-error (RMSE), and (c) correlation coefficient (CC). The ensemble mean bias, RMSE and CC are -0.06°, 0.54°C and 0.993, respectively.
Fig. 2. (left) Composite SST bias in the cross-track (x) and along-track (y) coordinates for all TCs (units: °C).
Fig. 3. (Right) Composite SST RMSE in the cross-track (x) and along-track (y) coordinates for all TCs (units: °C).
Bias
-
The mean bias of eight products is -0.06°C.
-
MW exhibits a bias of -0.26°C (highest).
-
MW_IR improves upon MW’s bias by 0.11°C, resulting in a bias of -0.15°C.
-
OISST and OSTIA performed relatively better than MW and MW_IR, with biases of -0.03° and -0.04°C, respectively.
-
NCEP FNL, ERA5 and JRA3Q show positive biases of 0.04°, 0.02°, and 0.11°C, respectively.
-
HYCOM exhibits a relatively large negative bias of -0.20°C.
-
The ensemble mean bias of these eight datasets is -0.06°C.
Root Mean Square Error (RMSE)
-
The mean RMSE of eight products is 0.54°C.
-
OISST shows the smallest RMSE at 0.37°C.
-
NCEP FNL exhibits the largest RMSE at 0.79°C.
-
MW and MW_IR have slightly higher RMSEs than OISST, at 0.56° and 0.49°C, respectively.
-
Similarly, the RMSE of OSTIA is 0.42°C
-
ERA5(atmospheric analysis-orreanalysis) shows an RMSE of 0.49°C, which is better than NCEP FNL by 0.30°C.
-
JRA3Q also outperforms NCEP FNL, with an RMSE of 0.61°C.
-
Additionally, HYCOM’s RMSE is close to JRA3Q’s.
Correlation Coefficient (CC)
-
The mean CC of eight products is 0.993.
-
OISST achieves the highest CC of 0.997.
-
OSTIA’s CC is slightly lower at 0.996.
-
MW and MW_IR have CCs of 0.994 and 0.995, respectively, which are weaker than those of OISST and OSTIA.
-
NCEP FNL exhibits the lowest CC at 0.985.
-
ERA5, JRA3Q and HYCOM have CCs of 0.994, 0.991 and 0.993, respectively.
The evaluation of bias and RMSE identifies OISST and OSTIA as the two most reliable models.
Furthermore, incorporating the CC metric confirms that OISST provides the most superior accuracy among all examined products.
2. Summary of Observation and Data Processing Methods for 8 Products
*Figure reference: Janjić, Tijana, et al (2018)
3. Cause Analysis of (+)/(-) Deviations in Product Data
3. 1. Causation of Cold Bias (-)
-
Satellites measure the extremely thin upper skin and sub-skin of the sea surface. Strong winds in the lower TC induce upwelling of low-temperature deep water; consequently, ocean models are often designed to overestimate the vertical mixing of the ocean in such cases.
-
In this case, intense precipitation and waves at the center of the TC generate noise in the brightness temperature(T_b; radiance emitted from the sea surface - to convert this observation to actual temperature, emissivity and atmospheric correction must be additionally calculated.), resulting in a cold bias in the SST calculation process.
3. 2. Causation of Warm Bias (+)
-
Atmospheric models generally use SST as a lower boundary condition. However, many atmospheric reanalysis datasets exhibit limitations in accurately representing the vertical mixing (3D) of the ocean in real time, resulting in its underestimation.
-
When employing a high-resolution 4D-Var data assimilation system such as ERA5, a high-quality background can be obtained by meticulously integrating various satellite data and field observations. A high-quality background also enables the most realistic simulation of the cooling effects induced by TC.
3. 3. Causation of Larger RMSE in Right-Rear Quadrant
-
Related to the mechanism of the TC: The TC moves along the edge of the North Pacific High due to beta-drift and steering flow. At this time, the right side of the TC engages with anticyclonic flow, (1) strengthening the pressure gradient force and (2) producing left-right asymmetry in wind speed. The strong winds on the right side of the TC induce Ekman pumping and vertical mixing, resulting in a Cold Wake, or sea surface cooling (SST cooling).
-
Due to the rapid temperature fluctuations and the model’s low resolution, the noise intensifies significantly. Accurately simulating asymmetric and rapid cooling at the precise location is challenging, and even a slight positional variation leads to a high RMSE.
4. Uncertainty caused by integration
Interpolation and multi-sensor integration are conducted to address gaps (such as clouds, orbital gaps, etc.) present in satellite data.
However, this process induces the following uncertainties:
(1) Limitations of Correlation Length: Since interpolation relies on surrounding data, smoothing errors arise in regions where temperature changes rapidly within a range of several kilometers, such as TC, resulting in a lower temperature gradient compared to the actual value.
(2) Inter-sensor bias: Different satellite sensors observe SST using distinct radiative mechanisms and sensing depths. Infrared (IR) sensors primarily measure the ocean skin temperature, whereas microwave (MW) sensors are sensitive to slightly deeper sub-skin temperatures. Although bias correction procedures are applied using reference sensors when constructing merged SST products, residual errors in the reference observations can propagate throughout the integrated dataset.
(3) Representation error: When integrating in-situ data, which are point observations, with satellite grid data that represent averages over areas spanning tens of kilometers, local variability within the grid is disregarded, leading to uncertainty.
While interpolation and multi-sensor integration are utilized to fill data gaps, they introduce critical uncertainties:
smoothing errors due to correlation length limitations, inter-sensor biases arising from varying measurement depths, and representation errors.
5.1. Integration of field observations with satellite data: OISST
Integration algorithm: Optimal Interpolation (OI)
1. Satellite deviation correction: Correct global deviations by comparing AVHRR infrared (IR) satellite data with in-situ measurements with buoys and vessels.
2. Conduct Optimal Interpolation on the corrected satellite data and field observations using a 0.25° grid to fill all grid gaps.
5.2. Integration of field observations with satellite data: OSTIA
Integration algorithm: Multi-scale OI
1. Hierarchical Integration: An interpolation method that progressively reduces errors across scales, from low-scale to high-scale,
such as ocean currents and fronts.
2. Data Source: Integrates IR and MW satellite data. (Relatively unaffected by the presence or absence of clouds.)
3. Foundation SST: Corrects by giving high weight to night time observations or applying diurnal cycle models to produce temperatures with diurnal cycles removed. Ultra-high resolution at 0.05°.
While OISST uses Optimal Interpolation (OI) to correct global satellite deviations using in-situ data on a 0.25° grid, OSTIA applies a multi-scale OI blending both IR and MW satellite data to achieve an ultra-high resolution of 0.05° with minimized cloud contamination.