version: "1.0.1" name: geoai description: >- Build browser geospatial AI apps with the geoai npm package (GeoAI.js). Use when integrating object detection, segmentation, classification, feature extraction, or image geolocation on satellite/aerial imagery; choosing a geoai task or map provider (esri, geobase, mapbox, tms, wms, oam, google); wiring geoai.pipeline / inference; web workers for background inference; or React + MapLibre patterns. Triggers include geoai, GeoAI.js, geobase AI, satellite detection, building footprints, DINOv3 embeddings, GeoCLIP, image geolocation, web worker, Worker API.

GeoAI.js

JavaScript/TypeScript library for running geospatial AI models in the browser (and Node). Imagery comes from map providers; inference runs via ONNX / Transformers.js.

Install

npm i geoai @huggingface/transformers onnxruntime-web

Peer deps are required. Package: `geoai`. Docs: https://docs.geobase.app/geoai — LLM index: https://docs.geobase.app/geoai/llms.txt

Core API

import { geoai } from 'geoai';
const pipeline = await geoai.pipeline(
  [{ task: 'object-detection' /* modelId?, modelParams? */ }],
  { provider: 'esri' } // see providers.md
);
const result = await pipeline.inference({
  inputs: { polygon: geoJsonPolygonFeature }, // GeoJSON Feature<Polygon>, required
  mapSourceParams: { zoomLevel: 18 }, // optional
  postProcessingParams: { confidence: 0.5 }, // task-specific, optional
});

• geoai.pipeline(tasks, providerParams) — one task or a chainable sequence.
• pipeline.inference(params) — returns detections / masks / features (task-dependent).
• geoai.models() — list registered tasks and metadata.

Polygon must be a GeoJSON Feature with Polygon geometry (not a bare geometry).

Zero-shot inputs

await pipeline.inference({
  inputs: {
    polygon: feature,
    classLabel: 'aircraft.wind turbine', // dot-separated classes
  },
  postProcessingParams: { threshold: 0.2, topk: 10 },
});

Task picker

Prefer specialized tasks over generic ones when they match.

Full task notes, defaults, and chain rules: tasks.md.

Providers

Configs: providers.md.

Background inference (Workers)

For UI apps, run the pipeline in a Web Worker so inference stays off the main thread. Vanilla JS and React patterns: workers.md.

React / MapLibre

There is no geoai/react export. Use the worker + hook pattern in react.md (builds on workers.md).

Agent checklist

1. Install geoai + peer dependencies.
2. Pick the most specific task id from the table above.
3. Configure a provider (start with esri if no credentials).
4. Pass a GeoJSON polygon Feature covering the AOI.
5. For UI apps, use a Worker so the main thread stays responsive (workers.md).
6. Read task docs for post-processing knobs before inventing params.

References

• tasks.md — registry tasks, models, chaining
• providers.md — provider param shapes
• workers.md — Web Worker API (vanilla JS + protocol)
• react.md — React hook + MapLibre UX
• Human docs: https://docs.geobase.app/geoai
• Live demos: https://docs.geobase.app/geoai-live
• Full LLM corpus: https://docs.geobase.app/geoai/llms-full.txt