Tessera (`tessera`)¶

Quick Facts¶

Field	Value
Model ID	`tessera`
Family / Source	GeoTessera precomputed embeddings
Adapter type	`precomputed`
Training alignment	N/A (precomputed product)

Tessera In 30 Seconds

Tessera is a precomputed 10 m global embedding product distributed as local GeoTessera tiles — rs-embed does no model inference here; it mosaics the tiles covering your ROI, reprojects the ROI into each tile's native CRS if needed, and returns the cropped (D,H,W) embedding grid in that tile-native CRS.

In rs-embed, its most important characteristics are:

strict north-up mosaic requiring consistent tile CRS/resolution — rotated/sheared or heterogeneous tiles raise: see Preprocessing / Retrieval Pipeline
year-selector temporal semantics: TemporalSpec.range(...) silently uses the start year rather than doing real temporal filtering: see Retrieval Contract
output CRS follows the tile-native CRS, not the EPSG:3857 default used by provider-backed paths elsewhere: see Output Semantics

Retrieval Contract¶

Field	Value
Backend	`auto` (legacy `local` still accepted)
`SpatialSpec`	`BBox` direct, or `PointBuffer` converted to EPSG:4326 BBox (approximate meter-to-degree)
`TemporalSpec`	`year(YYYY)` — uses `.year`; `range(start, end)` falls back to `start`'s year; default year `2021`
Source	GeoTessera precomputed tiles
Product CRS	tile-native (varies by tile), not EPSG:3857
Product resolution	10 m
Cache directory	`RS_EMBED_TESSERA_CACHE`, or per-call `sensor.collection="cache:/path/to/cache"`
Side inputs	none

Output CRS and temporal semantics

Tessera reads and returns embeddings in the product-native tile CRS after mosaic + crop. This differs from the EPSG:3857 default used by provider-backed models. TemporalSpec.range(...) is a year selector here, not scene-level temporal filtering.

Preprocessing / Retrieval Pipeline¶

flowchart LR
    INPUT["SpatialSpec\n+ TemporalSpec"] --> QUERY["ROI → EPSG:4326\n→ query tile blocks"]
    QUERY --> MOSAIC["Fetch tiles\n→ mosaic + crop"]
    MOSAIC --> POOL["pooled: vector"]
    MOSAIC --> GRID["grid: (D,H,W)\nin tile-native CRS"]

Architecture Concept¶

flowchart LR
    subgraph "Precomputed Tiled Product"
        RES["10 m per pixel"]
        STORE["GeoTessera tile blocks\n(local cache)"]
    end
    subgraph Retrieval
        RES --> BBOX["ROI → EPSG:4326 bbox"]
        STORE --> TILES["Query + fetch\ntile blocks"]
        BBOX --> TILES
        TILES --> MOS["Strict north-up\nmosaic + crop"]
    end
    subgraph Output
        MOS --> POOL["pooled: spatial\npooling over grid"]
        MOS --> GRID["grid: (D,H,W)\nin tile-native CRS"]
    end

Environment Variables / Tuning Knobs¶

Env var	Default	Effect
`RS_EMBED_TESSERA_CACHE`	unset (GeoTessera default)	Local GeoTessera cache directory
`RS_EMBED_TESSERA_BATCH_WORKERS`	`4`	Batch worker count for `get_embeddings_batch(...)`

Non-env override

sensor.collection="cache:/path/to/cache" overrides the cache directory for one call.

Output Semantics¶

pooled: spatial pooling over the cropped embedding grid.

grid: cropped (D,H,W) in product pixel space after mosaic + crop; metadata records input_crs=EPSG:4326 and output_crs follows the tile-native CRS.

Examples¶

Minimal example¶

from rs_embed import get_embedding, PointBuffer, TemporalSpec, OutputSpec

emb = get_embedding(
    "tessera",
    spatial=PointBuffer(lon=121.5, lat=31.2, buffer_m=5000),
    temporal=TemporalSpec.year(2021),
    output=OutputSpec.pooled(),
    backend="auto",
)

Example cache override¶

# Example (shell):
export RS_EMBED_TESSERA_CACHE=/data/geotessera

Paper & Links¶

Publication: CVPR 2026

Reference¶

Tile mosaic requires all tiles to be north-up with consistent CRS and resolution — tiles with rotation/shear are rejected.
Output CRS is tile-native (not EPSG:3857) — do not compare grids with provider-backed models without reprojecting.
"No tiles found" usually means the ROI/year combination has no coverage in the GeoTessera cache, not that the cache is broken.

Tessera (tessera)¶