Gridded sea ice concentration (SIC) data are fundamental for understanding long-term polar climate. Since 1978, daily SIC data have been derived from passive microwave (PM) satellite observations. As SIC is not directly observed, it is obtained through an algorithmic inversion of PM brightness temperatures. Differences between PM-SIC algorithms lead to uncertainty in SIC data. Here we examine differences between several PM-SIC products and focus primarily on two: the National Snow and Ice Data Center (NSIDC) Climate Data Record (CDR), which over our study period is derived from Special Sensor Microwave Imager/Sounder (SSMIS) data, and over the Southern Ocean is largely based on the Bootstrap (BT) PM-SIC algorithm, and the Advanced Microwave Scanning Radiometer 2 (AMSR2)-based NASATeam2 (NT2) product. On individual days, changes in integrated Antarctic sea ice area (SIA) can reach as high as 1.6 million km2 in summer with relative differences up to 37% at the annual sea ice minimum. Using more than 1.8 million coincident observations of the ICESat-2 Linear Ice Fraction (LIF) from 2018 to 2024, we find BT-based PM-SIC data underestimate Antarctic SIC by a median of 0.2 in areas classified as the marginal ice zone (MIZ), corresponding to a median relative difference in SIC of 31%. This offset is seasonally, spatially, and parametrically consistent and yields differences in apparent MIZ extent as high as 4.4 million km2. The offset exceeds reported CDR uncertainties that average 0.07 and cannot be explained as a consequence of sensor differences. The parametric simplicity of the observed bias means these offsets can be parameterized as a function of BT SIC alone, which could lead to an improved SIC dataset for Antarctic studies.
Landsat has provided an indispensable foundation for field-scale evapotranspiration (ET) mapping but is constrained by its overpass temporal frequency. Although integrating additional thermal data offers a pathway to bridge this gap, the magnitude of improvement varies across observation characteristics and surface context and remains less understood. Here, ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) is used as an additional thermal source on top of the Landsat baseline to evaluate the value of more frequent thermal observations for improving field-scale ET time series. We implemented a Google Earth Engine -based Disaggregated Atmosphere-Land Exchange Inverse + Data Mining Sharpener framework to generate a Landsat-based ET baseline and enhanced it with ECOSTRESS observations. The resulting ET data was validated across 25 flux-tower sites spanning diverse land-cover types and climate zones across the contiguous United States. When integrated with Landsat, the addition of ECOSTRESS observations reduced interpolation errors and improved temporal consistency across most sites. However, improvements in ET time series agreement depended strongly on (1) the extent to which ECOSTRESS reduced effective Landsat sampling gaps and (2) the accuracy of overpass-day ECOSTRESS ET retrievals. Improvements were largest when ECOSTRESS shortened Landsat observation gaps by ~50% and maintained high scene quality. ECOSTRESS overpass time also influenced performance, with acquisitions between 10:00–14:00 producing most reliable ET estimates. These results support developing methods with multi-satellite observations to strengthen field-scale ET and help address the key knowledge gap regarding the value of increased thermal observations.
Nitrogen dioxide (NO2) is a major air pollutant with significant impacts on human health, yet satellite-based estimates of surface NO2 with full diurnal coverage remain limited. Here, we develop a machine learning framework to estimate hourly near-surface NO2 concentrations across China by integrating complementary observations from two Chinese polar-orbiting satellites, GF-5B (~10:30 local time) and DQ-1 (~13:30 local time). A Light Gradient Boosting Machine (LightGBM) model driven by twin-satellite NO2 vertical column densities (VCDs) and meteorological reanalysis achieves strong predictive performance in 10-fold cross-validation, with a coefficient of determination (R2) of 0.91 and root mean square error (RMSE) of 5.40 μg/m3 across all hours (4.46 μg/m3 during daytime; 6.32 μg/m3 at night), and retains reasonable skill under stricter site-based, time-based, and independent cross-year validation. The reconstructed dataset captures the characteristic diurnal cycle of NO2, with morning and evening peaks associated with traffic emissions and a midday minimum driven by photochemical loss. Compared to the single-satellite TROPOMI product, the twin-satellite configuration substantially improves the representation of diurnal variability, increasing R2 from 0.81 to 0.91 and reducing RMSE by 31.60%, and corrects a distorted urban-rural NO2 exposure ratio (1.70 for TROPOMI versus a ground-based 1.19), narrowing it to 1.30. Feature-importance analysis further shows that meteorological variables dominate model predictions during both day and night, with satellite VCDs providing a complementary, though indirect, constraint on nighttime concentrations. The resulting hourly NO2 dataset demonstrates the value of synergistic multi-satellite observations for air quality monitoring and provides a robust basis for refined pollution control and population exposure assessment.
