ETHERIUM Workspace

Loading map data

Area of interest

Zoom past level 13 to pick a block

—

Date

Search window ± 90 days

Reading the active date from the tile engine.

Analysis

Index

Export

Enhance a file

Drop a GeoTIFF, or choose one Up to 1536 px on the long side

Quick look

Read from the enhanced RGB picture, not from Sentinel bands. Use it to see a tile quickly; use the Index tools above for any number you intend to defend.

Select a site on the map to read one of its tiles.

Ctrl-click a block past zoom 13 to open it here, and compare the Sentinel-2 input with the model output for that tile.

Provenance

Zonal statistics

Run an analysis to fill this table. One row per region, with the index mean taken on the sensor's own grid.

Model report

Tested on 7 places in India the model never saw in training, 2,800 tiles. Input is 9.55 m Sentinel-2 L2A (8 dates stacked into 24 channels); output is 2.39 m pixel spacing, 4x more pixels in each direction. Pixel spacing is not resolution: resolvable detail is coarser.

In season

Sentinel-2 dates close to the aerial reference image.

Model	cPSNR (dB)	SSIM	LPIPS	Edge-F1
Bicubic upscaling	20.58	-	0.588	0.405
India model (starting point)	20.10	0.427	0.233	0.641
Model 1 (B)	20.39	0.459	0.197	0.642

Off season, about 6 months from the reference: Model 1 LPIPS 0.233 vs 0.266 for the India model and 0.611 for bicubic; Edge-F1 0.570 vs 0.565 vs 0.350.

Weak spots: snow and flat farmland are hardest (snow LPIPS about 0.31), and the models don't beat bicubic on cPSNR, a metric that rewards blur.

What the numbers mean

cPSNR: Pixel-by-pixel agreement with the aerial image, forgiving small shifts and brightness offsets.
SSIM: Structural similarity, from 0 to 1.
LPIPS: How different two images look to a network trained to mimic human judgement. Lower is better.
Edge-F1: Whether roads, building outlines and field boundaries are in the same places as in the aerial image.

Models

Model 1: ArcGIS colours: Trained against ArcGIS World Imagery, so it looks like the aerial map. Variant B was chosen.
Model 2: Sentinel colours, computed: Model 1's detail, recoloured so every 4x4 block matches the Sentinel-2 input's colour.
Model 3: Sentinel colours, trained: Learns the same goal directly. Training in progress; no results yet.

Use it carefully

The output is a prediction, not real 2 m imagery.
Index values are Sentinel-2 native resolution and are never super-resolved.
Boundaries are model-derived and may be wrong where confidence is low.
Check it before anything safety-critical or legal.
Trained and tested on India only: 55 places, about 104,000 training pairs.