Geospatial vision + vision-language foundation models

The GeoFM landscape, mapped.

Models, a learning roadmap, open problems and benchmarks for Earth-observation foundation models and RS-VLMs.
All Production-ready Frontier (2025–26) Foundational Both tracks Vision FM Vision-Language SAR Multispectral Temporal Text/VLM

Learning roadmap

Nine phases, about 10–16 weeks part-time: the vision track, then the vision-language track (start with the Anatomy explainer), then convergence. Ticks are saved in this browser. Full detail in 01-roadmap/roadmap.md.

Next-gen gap map

14 open problems from GEO-Bench-2, PANGAEA, RSMMVP and the 2025–26 literature (evidence refreshed Nov 2026). Where two or three gaps meet, there's a design candidate. Working file: 04-next-gen/gap-map.md.

Benchmarks

Seed data compiled from papers; verify before citing and extend via 03-registry/leaderboard.json. The adaptation protocol changes who wins, so read every leaderboard with its protocol attached. Eval tooling: GEO-Bench · PANGAEA · TerraTorch · GEOBench-VLM · VLMEvalKit · lmms-eval.

All Vision FM VLM Agents

Workflow

How this repo is meant to be used, from learning to a prototype.

1

Learn

Work through the roadmap phases. Vision track (1–5) → VLM track (5b–5d) → convergence (6). Start VLM concepts at the Anatomy page.

2

Dissect papers

Per key paper: copy 02-papers/TEMPLATE.md, fill the 5 layers. For VLMs, also fill the model's row in the Comparison Matrix — predict the three slots from the abstract, then correct after reading.

3

Track landscape

New model drops → add to 03-registry/models.json → this page updates automatically (single source of truth). Ask Claude for a weekly arXiv/HF scan.

4

Map gaps

Each dissection's "next-gen implications" feeds 04-next-gen/gap-map.md. Look for intersecting gaps — the convergence thesis (G10+G11+G12) lives there.

5

Design

Draft an architecture spec targeting 2–3 gaps. Design review with Claude's rs-dl-architect skill: tokenization, objective, eval plan.

6

Prototype & evaluate

Smallest viable version. Vision: GEO-Bench-2 subset, linear probe + LoRA. VLM: LHRS-Bench + RSMMVP probes + grounding IoU, against a frontier generalist baseline.

Evaluation discipline: linear-probe before fine-tune · few-shot scaling at 100/1k/10k labels · cross-region hold-outs · DINOv3 baseline for vision, GPT-4o-class baseline for VLMs · report grounding IoU separately from text metrics · probe hallucination explicitly · for reasoning models, check answer–grounding consistency (AESP), not just accuracy.