Earth observation field notes
Field notes on large-area Earth observation — the resolution, cadence and processing choices that decide whether a satellite product is actually usable, and where EarthToDate fits.
Worldwide air quality from orbit: Sentinel-5P, absolute and year-on-year
Sentinel-5P measures NO₂, SO₂, CO, CH₄, HCHO and aerosols daily, worldwide. Why the year-on-year delta against a same-period baseline beats a single snapshot for spotting real change.
Read →GLOF early warning: model the water that actually flows
Most glacial-lake risk assessments report total lake volume — the wrong number. What determines downstream impact is the volume that would actually discharge if the moraine fails, and that needs a metre-accurate DEM.
Read →Why pansharpening fails on some satellites — and how to fix it
Panchromatic and multispectral bands aren't always co-registered. On off-nadir RPC products the offset varies across the scene, producing colour fringing that a single shift can't remove. A measure-then-align approach.
Read →Flood mapping is a terrain problem — why 30 m DEMs fail
Answering 'what floods if water reaches this height' depends entirely on elevation accuracy. Public 30 m DEMs blur out the kerbs, embankments and drainage lines that decide where water actually goes.
Read →Land cover without training data: a physics-based pipeline
Supervised land-cover classifiers need labels and generalise poorly across geographies. A training-free approach clusters materials by spectrum and reads spatial context, so built-up, roads and cropland emerge as queries against one representation.
Read →Finding dark ships: AIS over SAR
AIS tells you who is reporting; SAR shows every hull, day or night, through cloud. The intelligence is in the mismatch — how fusing the two surfaces dark ships and AIS spoofing.
Read →Detecting unpermitted construction from dated imagery
Unpermitted construction is small, dispersed and constant. Detecting pixels that are built now but weren't months ago — at national scale — needs a false-positive discipline, not just a classifier.
Read →Mining monitoring: 2 m weekly worldwide vs 480 m quarterly
Free mining-detection data exists — but it's Amazon-only, quarterly and 480 m. Why resolution and cadence decide how many sites you actually find, and what typed detection adds for enforcement.
Read →Catching crop problems at 4×4 m — weeks earlier
A native Sentinel-2 pixel is 100 m²; a problem must dominate it to register. At 4×4 m, divergence from the local trend surfaces disease, pests, weeds and irrigation faults before they cost a season.
Read →After WorldView-3: a 2–4 m SWIR path for mineral mapping
WorldView-3 retired in December 2025, ending the only sub-5 m shortwave-infrared source for hydrothermal alteration mapping. How enhanced Sentinel-2 fills the gap at 2–4 m — worldwide and without tasking.
Read →A 1 m global elevation model, within a metre of LiDAR
Public DEMs are 30 m with 5–16 m vertical error; survey LiDAR is episodic and local. A consistent 1 m global DEM changes flood, drainage, drill-targeting and earthworks work.
Read →Country-scale water-leak detection from Sentinel-2
Utilities lose 20–40% of treated water. Tasking-SAR services can't cover a national network affordably. How 2 m soil-moisture indices from Sentinel-2 change the unit economics.
Read →Physics-based enhancement vs hallucinated super-resolution
Generative super-resolution invents plausible detail. For measurement, geolocation and change detection, invented pixels are a liability. What 'enhancement without hallucination' means in practice.
Read →