ASI Prism — physics-based spatial, spectral and multimodal inference for Sentinel-2

ASI Prism — physics-based spatial, spectral and multimodal inference for Sentinel-2. Type coordinates or click the map, then Load Sentinel-2: the least-cloudy scene in your date range is pulled straight from Microsoft Planetary Computer (no login), windowed to your spot, and dropped into the same live-inference cockpit — method / knob α / band display / wipe comparison / trust gate / results table all work exactly as before. Every load also lists the other dates over that spot (switch to compare, e.g. before/after an event). Needs internet; imagery © ESA/Copernicus via Planetary Computer.
1 · Pick a location 2 · Load & render the scene 3 · Render results 4 · Wipe comparison 5 · Extract every head nothing has run yet — start at step 1
1Location
Drag the map's bottom edge to change its height; drag any figure's bottom-right corner to resize every figure on the page.
Click the map or type coordinates, then Load Sentinel-2.
2Load & render this scene
SpecErr lowhigh — per-pixel SAM(down(SR), LR), lower is better. "multiscale x2/x4/x8" are ONE model rendered at different target GSDs; benchmark HA/OM/IM metrics apply at native x4 only. "NIR (true band)" shows the real near-infrared channel; grayscale/R/G/B act on the rendered display. Figure size: drag or +/- keys.
3Render results for the loaded scene — upsampled input, reconstruction, consistency error, bands
LR input / bicubic
SR
HR / SpecErr
4Advanced wipe comparison — any layer / method / α on each side
Drag the divider. LR and LR (raw) hold exactly the same measurements — the only difference is how they are drawn. LR (raw) keeps a hard step at every 10 m pixel boundary, and hard steps read as sharpness, so it can look crisper while containing no more detail. LR is that same data interpolated, and is the standard baseline a reconstruction should be judged against: it is the free alternative to running a model, and the stronger of the two to beat. Band display applies here too.
5Extract — tick what you want from this spot, then run once. Everything works from the one loaded Sentinel-2 scene, except land-surface temperature, which also fetches its own Landsat scene. Results arrive as tabs below.
provenance manifest (sources, terms, checkpoint digests, what is not cleared for public showing)




6Individual heads (advanced) — run one at a time, same models, no tabs

 

  Cross-modal prediction
  — one loaded optical scene in, a different sensor or interpretation out, one head at a time. Every output below is a PREDICTION (regime R3): no radar, thermal or land-cover instrument measured this spot. Results appear under the buttons.
 

 
Sentinel-1 radarOptical reflectance in → VV/VH backscatter out. No radar is measured here.

  

 
Land-surface temperatureFetches a separate Landsat-8/9 scene, then shows Landsat's own measured LST beside the predicted one, so the error map is real.

  

 
Land coverOptical in → per-class posteriors at native 10 m. An interpretation, not a measurement.

  

 
Relative reliefOptical in → relief in metres. Weak v1: in-domain Pearson r = 0.0372 and undefined off-domain. A wiring demo.

  

 

 

 
SAR predicts Sentinel-1 backscatter from the loaded optical crop (pick asc/desc). Thermal fetches a Landsat-8/9 scene for these coordinates and predicts land-surface temperature, shown next to Landsat's own measured LST so the error map is real. Land cover predicts classes for the selected taxonomy — WorldCover 8-class (default) or GLC_FCS10 30-class (RESEARCH: richer but ~half its classes are data-starved; trust the well-supported classes + the coarse super-class coverage) — R3 interpretation, native 10 m, uncalibrated max-sigmoid posterior. DEM relative-relief is an EXPERIMENTAL weak v1 (in-domain corr ~0.04) — a prediction, not a product capability.







 

  Rendered results — this scene
  
  
  Rows come from the disk cache; click headers to sort, click a row to load it.