The satellite maps got better by admitting their errors
A widely used elevation dataset carries metre-scale tile errors, coal-mine maps suffer when extra bands are added, and a new altimeter reads tides in a small Arctic fjord. All three are preprints.
- GEM-15 contains systematic tile offsets of -4.4 to +5.0 m and periodic stripes of ±0.6 m. Self-reported m:…9872
- Slope-adaptive fusion of GEM-15 and Copernicus GLO-30 reduces RMSE by 21% in Saudi Arabia and 30% in the USA. Self-reported m:…9872
- Increasing spectral bands in ImageNet-pretrained SegFormer models leads to spectral dilution, reducing mine IoU to 0.0094 at 15 bands. Self-reported m:…2154
- SWOT data shows an overall correlation of r=0.91 with CHS harmonic tide predictions in the Sugluk Inlet. Self-reported m:…4668
The elevation map with seams
A preprint on a widely used elevation dataset starts with a diagnosis. GEM-15 contains systematic tile offsets of -4.4 to +5.0 m and periodic stripes of ±0.6 m, and the stripes are inherited from the ALOS PRISM reference grid with spacing of 47-53 km. The errors are not random noise. They are structure left over from the reference the dataset was built on, repeated tile after tile. m:…9872
The errors have consequences on flat ground. Uncorrected GEM-15 tile seams can displace water lines by 0.3-1.1 km on flat plains, which is enough to put a flood boundary in the wrong place. The paper offers two repairs: orbit-guided destriping reduces GEM-15 NMAD by 9-12% against lidar and ICESat-2, and slope-adaptive fusion of GEM-15 and Copernicus GLO-30 reduces RMSE by 21% in Saudi Arabia and 30% in the USA. The dataset got better not by collecting new data, but by admitting what was wrong with the old data. m:…9872
More bands, worse maps
A second preprint, on mapping surface coal mines from Sentinel-2 satellite imagery, finds the opposite of what more data usually promises. Prithvi-EO-2.0 achieves superior performance in surface coal mine segmentation compared to traditional architectures when using multi-spectral Sentinel-2 imagery, but increasing spectral bands in ImageNet-pretrained SegFormer models leads to spectral dilution, reducing mine IoU to 0.0094 at 15 bands. Adding bands made the traditional model dramatically worse, not better. m:…2154
The stakes are not cartographic. Current global methane emission inventories for coal mines are limited by kilometer-scale spatial uncertainty due to reliance on point coordinates, which means the inventories know roughly where the mines are but not exactly. A better map is a better inventory, and a better inventory is the difference between attributing emissions to the right mine and the wrong one. m:…2154
Tides in a fjord
A third preprint tests whether the SWOT satellite can read tides where it should struggle: a small Arctic fjord. SWOT KaRIn altimetry, in its Level-2 Low-Rate form, can detect tidal water-level variations in small Arctic fjords, and SWOT data shows an overall correlation of r=0.91 with CHS harmonic tide predictions in the Sugluk Inlet. A correlation of 0.91 against an independent tide prediction is a strong result for a place satellites were not designed to measure. m:…4668
The paper is honest about the limit. SWOT recovers nearly full tidal amplitude over open straits, with a slope of 0.9, but shows reduced amplitude in narrow inner inlets. The instrument works where the water is open and weakens where the fjord narrows, which is exactly the kind of boundary a user needs to know before trusting the numbers. m:…4668
Error budgets as progress
The three papers share a habit that Earth observation is learning late: publish the error budget alongside the result. A fused elevation map that is 30% more accurate in the USA, a mine map that gets worse when bands are added, and a tide measurement that works in the open strait but not the narrow inlet are all more useful than a clean-looking map with hidden flaws. All three are preprints, so the numbers still await peer review. But the direction is right: a map that says where it is wrong is worth more than a map that pretends it is not. m:…9872 m:…2154 m:…4668
What is still unresolved
- All three results come from preprints; none has been peer-reviewed.
- Whether the GEM-15 destriping and fusion methods generalize beyond the tested regions is not shown in the cited claims.
- Whether Prithvi-EO-2.0 holds its advantage on mines outside the benchmark dataset is untested.
- Whether SWOT tide measurements work in other Arctic fjords beyond the Sugluk Inlet is unknown.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.