Geospatial Tools plugin

mere-geo-tools is the portable workflow layer around core mere.run geo models. It prepares exact source imagery, preserves raster and preprocessing provenance, invokes native Swift/MLX inference, and georeferences the result.

Node	Inputs	Outputs
geo.flood.segment	Four S2 L2A, S1 RTC, and DEM timesteps	Candidate mask, probability COG, preview, manifest
geo.fire.segment	Four S2 L2A, S1 RTC, and DEM timesteps	Candidate mask, probability COG, preview, manifest
geo.tessera.embed	Annual S2 plus ascending or descending S1 histories	Per-pixel embedding safetensors and manifest
geo.olmoearth.embed	One to twelve S2, S1, or Landsat observations	Per-modality spatial embedding safetensors and manifest

Install and inspect

mere.run plugin install mere-geo-tools --yes
mere-geo-tools manifest --json
mere-geo-tools doctor --json
mere-geo-tools graph catalog --json
mere-graph-conformance --provider mere-geo-tools --json

Prepare pinned inputs

All recipes declare WGS84 bounds and the target WGS84 UTM grid. Sources are exact STAC collection/item pairs; preparation never silently searches for a different scene.

Hazard recipes use four ordered roles and differ only in their typed kind:

{
  "kind": "mere.geo/terramind-fire-source-recipe",
  "version": 1,
  "sample_id": "fire-review-001",
  "target": {"aoi": [-123.3, 49.1, -123.2, 49.2], "crs": "EPSG:32610"},
  "timesteps": [
    {"role": "pre_month", "S2L2A": {"collection": "sentinel-2-l2a", "item": "..."}, "S1RTC": {"collection": "sentinel-1-rtc", "item": "..."}},
    {"role": "pre_event", "S2L2A": {"collection": "sentinel-2-l2a", "item": "..."}, "S1RTC": {"collection": "sentinel-1-rtc", "item": "..."}},
    {"role": "event", "S2L2A": {"collection": "sentinel-2-l2a", "item": "..."}, "S1RTC": {"collection": "sentinel-1-rtc", "item": "..."}},
    {"role": "post_event", "S2L2A": {"collection": "sentinel-2-l2a", "item": "..."}, "S1RTC": {"collection": "sentinel-1-rtc", "item": "..."}}
  ],
  "DEM": {"collection": "cop-dem-glo-30", "item": "..."}
}

TESSERA keeps independent S2, ascending S1, and descending S1 timelines:

{
  "kind": "mere.geo/tessera-v2-source-recipe",
  "version": 1,
  "sample_id": "annual-context-001",
  "target": {"aoi": [-123.3, 49.1, -123.2, 49.2], "crs": "EPSG:32610"},
  "observations": {
    "S2": [{"collection": "sentinel-2-l2a", "item": "..."}],
    "S1_ASC": [{"collection": "sentinel-1-rtc", "item": "..."}],
    "S1_DESC": [{"collection": "sentinel-1-rtc", "item": "..."}]
  }
}

OlmoEarth uses a shared timeline; every timestep must contain the same selected modalities:

{
  "kind": "mere.geo/olmoearth-v1.2-source-recipe",
  "version": 1,
  "sample_id": "multisensor-context-001",
  "target": {"aoi": [-123.3, 49.1, -123.2, 49.2], "crs": "EPSG:32610"},
  "timesteps": [
    {
      "observed_at": "2026-06-15T10:00:00Z",
      "S2L2A": {"collection": "sentinel-2-l2a", "item": "..."},
      "S1RTC": {"collection": "sentinel-1-rtc", "item": "..."}
    }
  ]
}

Sentinel-2 L2A and Sentinel-1 RTC recipes use their canonical STAC asset names and are exercised against the Planetary Computer catalog. Landsat is deliberately fail-closed: OlmoEarth expects raw unsigned 16-bit OLI/TIRS Level-1 DN values in the canonical B8, B1, B2, B3, B4, B5, B6, B7, B9, B10, B11 order. Every Landsat item must declare source_contract as landsat-oli-tirs-level1-dn-v1 and provide an explicit assets map for all eleven bands. Planetary Computer's landsat-c2-l2 collection is not compatible because it lacks the required tensor and has different radiometry.

Prepare and inspect any recipe with the same commands:

mere-geo-tools prepare --recipe ./sources.json --output ./prepared-inputs --json
mere-geo-tools inspect ./prepared-inputs --json

Each bundle records source item/asset identities, acquisition times, canonical band order, cloud-mask or radar conversion policy, grid metadata, and SHA-256 hashes for every materialized input.

Hardware scaling

Leaving model as auto delegates tier selection to core mere.run based on unified memory and installed checkpoints. Graphs may explicitly select every managed tier, including TESSERA Teacher and OlmoEarth Base. TESSERA's batch_pixels can be increased on larger machines. OlmoEarth exposes patch_size, input_resolution, and include_tokens; smaller patches and full tokens preserve more detail and consume more memory.

The neural commands are mere.run geo flood, geo fire, geo tessera, and geo olmoearth. There is no PyTorch fallback in this provider.

Evidence and license boundaries

Flood and fire masks remain candidate-only until independently corroborated and reviewed. TESSERA and OlmoEarth outputs are derived-feature artifacts, not findings. OlmoEarth use remains subject to its upstream artifact license and its restrictions on military, defense, intelligence, human-surveillance, policing, and listed extractive uses.