Skip to content

Supported Models (Overview)

This page is the model selection entry point.

Use it to quickly narrow down candidates, then jump to:

How To Use This Page

Recommended flow:

  1. Pick a shortlist from the overview tables below
  2. Validate temporal + input assumptions in Advanced Reference
  3. For production/benchmark runs, confirm model-specific caveats and env vars before exporting

Quick Chooser by Goal

Goal Good starting models Why
Fast baseline / simple pipeline tessera, gse, copernicus Precomputed embeddings, fewer runtime dependencies
Simple S2 RGB on-the-fly experiments remoteclip, satmae, scalemae, dynamicvis Straightforward RGB input paths
Time-series temporal modeling agrifm, anysat, galileo Native multi-frame temporal packaging
Multispectral / strict spectral semantics dofa, terramind, thor, satvision Strong channel/schema assumptions
Mixed-modality experiments (S1/S2) terrafm Supports S2 or S1 path (per call)

Available Detail Pages (Current)

Canonical model IDs are the short names shown below. Legacy IDs (for example remoteclip_s2rgb, prithvi_eo_v2_s2_6b, thor_1_0_base) are still accepted as aliases in the API/CLI.

More model detail pages can be added using the Model Detail Template.

Model Catalog Snapshot

Precomputed Embeddings

Model ID Type Primary Input / Source Outputs Temporal mode Notes Detail
tessera Precomputed GeoTessera embedding tiles pooled, grid yearly coverage product Fast baseline, source-fixed precomputed workflow detail
gse Precomputed Google Satellite Embedding (annual) pooled, grid TemporalSpec.year(...) Annual product via provider path detail
copernicus Precomputed Copernicus embeddings pooled, grid limited (2021) Coarse resolution product detail

On-the-fly Foundation Models

Model ID Primary Input Temporal style Outputs Notable requirements Detail
remoteclip S2 RGB (B4,B3,B2) single composite window pooled, grid provider backend; RGB preprocessing detail
satmae S2 RGB (B4,B3,B2) single composite window pooled, grid RGB path; ViT token/grid behavior detail
scalemae S2 RGB + scale single composite window pooled, grid requires sensor.scale_m / input_res_m detail
dynamicvis S2 RGB (B4,B3,B2) single composite window pooled, grid larger default resize (512) detail
wildsat S2 RGB single composite window pooled, grid normalization options detail
prithvi S2 6-band single composite window pooled, grid required temporal + location side inputs detail
terrafm S2 12-band or S1 VV/VH single composite window pooled, grid choose modality per call detail
terramind S2 SR 12-band single composite window pooled, grid strict normalization/channel semantics detail
dofa Multispectral + wavelengths single composite window pooled, grid wavelength vector required/inferred detail
fomo S2 12-band single composite window pooled, grid normalization mode choices detail
thor S2 SR 10-band single composite window pooled, grid strict stats-based normalization detail
satvision TOA 14-channel single composite window pooled, grid strict channel order + calibration detail
anysat S2 10-band time series multi-frame pooled, grid frame dates (s2_dates) side input detail
galileo S2 10-band time series multi-frame pooled, grid month tokens + grouped tensors detail
agrifm S2 10-band time series multi-frame pooled, grid fixed T frame stack behavior detail

Temporal and Comparison Notes (What People Usually Miss)

  • TemporalSpec.range(start, end) is usually a window for compositing, not a single-scene acquisition selector.
  • OutputSpec.grid() may be a token/patch grid, not a georeferenced raster grid (especially for ViT-like backbones).
  • Cross-model comparisons are easiest with OutputSpec.pooled() and fixed ROI/temporal/compositing settings.
  • Multi-frame models (agrifm, anysat, galileo) need extra attention to frame count and temporal side inputs.

Read the details in Supported Models (Advanced Reference).

Per-model Detail Pages (Template-first)

This repo now includes a reusable template page for documenting each model consistently:

Recommended fields for each model page:

  • what the model expects (input contract, band order, temporal mode)
  • what rs-embed currently feeds (current adapter behavior)
  • preprocessing defaults and env knobs
  • output semantics (pooled vs grid details)
  • caveats for reproducibility / fair benchmarking

Source of Truth in Code

Model registration source of truth:

  • src/rs_embed/embedders/catalog.py (MODEL_SPECS)

Implementation details for each adapter live in:

  • src/rs_embed/embedders/onthefly_*.py
  • src/rs_embed/embedders/precomputed_*.py

When docs and code disagree, check code first and update docs accordingly.