The satellites watching Earth keep arriving with their own error bars

What free images cannot see

The most useful finding in this week's Earth observation papers is a negative one. A preprint testing free Sentinel-2 satellite imagery against Himalayan rock-ice collapses found no distinctive pre-collapse precursors at three major failure sites: Chamoli, Langtang Lirung and Budhi Gandaki. The creep-mapping method the author tested can identify steady ice movement, but it cannot serve as a predictive early-warning system for catastrophic collapses. What a free-data screen offers instead is a coverage map and a candidate list of slopes to watch, not a warning. m:…0032

The limits are starkly altitude-dependent. Detection of slope creep falls from 99% at 4,000 metres to 35% at 5,500 to 5,750 metres, and to 10-17% above 6,000 metres. That matters because the highest, most dangerous slopes are exactly where the free imagery goes blind. For infrastructure teams and risk analysts, the paper draws a clean line: open-access optical imagery is a screening tool with a known ceiling, and anyone treating it as an early-warning system is overclaiming what the data can do. m:…0032

Where the new instruments work

The SWOT satellite does considerably better where it was designed to look. A preprint processing SWOT data across the river-estuary-ocean continuum in large deltas produced continuous water surface elevation profiles validated against ground gauges, with errors of 0.32 metres in the Mekong delta and 0.34 metres in the Meghna. That precision is enough to quantify how tides and river discharge interact: discharge is the main control on how fast tidal troughs travel upstream and how far they penetrate, while tidal amplitude shapes how sharp those troughs are. For flood modellers, that turns a satellite pass into a measurement of compound-flood risk in deltas where ground gauges are sparse. m:…0031

TEMPO, the geostationary air-quality instrument, passed a similar cross-check. Compared against the established TROPOMI and OMI instruments, its nitrogen dioxide and formaldehyde products correlated strongly once pixels were aggregated to 0.2 degrees: nitrogen dioxide correlation rose from 0.58-0.70 to 0.91-0.96, and formaldehyde from 0.34-0.36 to 0.71-0.89. TEMPO also resolved localised pollution enhancements around cities, power plants and fires. The aggregation step is the honest part: at native resolution the agreement is weaker, and the paper shows exactly how much smoothing the comparison needs. m:…6480

The old instruments were wrong too

A Greenland study shows the problem is not only with new satellites. High-fidelity surface observations from the southwestern Greenland ice sheet found that optically thin clouds increase absorbed surface radiation by 26.5 watts per square metre on average compared with clear conditions, and near solar noon the peak under cloudy skies was 80% higher, 63 against 35 watts per square metre. Standard automatic weather stations underestimate the daily mean cloud effect by 30% and overestimate the peak clear-sky flux by a factor of two. Long-term sea-level projections built on those station records carry that bias with them. m:…3063

The same candour runs through a preprint that reconstructs the global ocean from 1950 to 2020 using a diffusion model trained on climate simulations and conditioned on sparse historical observations. The reconstruction recovers large-scale upper-ocean features and matches satellite altimetry with a median correlation of 0.41 for sea-surface height, and basin-scale heat content aligns with observations after 1990. But it carries a warm offset in heat content before 1990, most pronounced in the South Pacific, its ensemble is underdispersed, and it fails to capture the Antarctic sea-ice decline. The authors report the misses alongside the hits, which is what makes the reconstruction usable: a risk modeller can see where to trust it and where not to. m:…0030

The pattern

Five papers, one working method. State what the instrument can resolve, state the error, and state where it stops working. Sentinel-2 screens slopes but cannot warn of collapses, and goes blind above 6,000 metres. SWOT measures delta water levels to a third of a metre. TEMPO needs its pixels aggregated before it agrees with its predecessors. Greenland's weather stations misread the cloud effect by 30%. The ocean reconstruction works after 1990 and drifts before it. None of this is a reason to distrust satellite Earth observation. It is the reason to trust it: the field is now reporting its limits as carefully as its results, and the limits are where the next instrument, and the next decision, should start. m:…0032 m:…0031 m:…0030