Skip to content

index.md

Latest commit

 

History

History
63 lines (52 loc) · 3.61 KB

index.md

File metadata and controls

63 lines (52 loc) · 3.61 KB

TEE Documentation

Tessera Embeddings Explorer (TEE) is a Django + vanilla JavaScript web application for interactive exploration of satellite embedding vectors produced by the Tessera foundation model (accessed via the GeoTessera Python library). Users define geographic viewports, browse PCA/UMAP projections of 128-dimensional embeddings, create land-cover labels via similarity search, run change-detection across years, and evaluate classifiers against ground-truth shapefiles.

Table of Contents

Document Description
architecture.md System architecture, panel layout, module graph, state management, dependencies, testing, deployment
frontend_api.md JavaScript API reference for all 8 ES modules, data structures, frontend→backend call map
backend_api.md Python HTTP endpoints, library modules, standalone scripts
extension_guide.md Recipes for adding panels, classifiers, data sources, endpoints, label types — with checklist

Key Features

  • Six-panel viewer with four modes: explore, change-detection, labelling, validation
  • Client-side similarity search over 128-dim float32 embeddings (~250K vectors per viewport) using brute-force L2
  • PCA and UMAP dimensionality reduction rendered as interactive 3D scatter plots (Three.js)
  • K-means segmentation via inline Web Worker with K-means++ initialisation
  • Change-detection heatmap comparing per-pixel embedding distances across years
  • Manual labelling with point, polygon, and similarity-expansion label types
  • Hierarchical schema browser for structured label ontologies (UKHab v2, HOTW, EUNIS, or custom)
  • ML evaluation pipeline with streaming NDJSON learning curves (k-NN, RF, XGBoost, MLP, spatial MLP, U-Net)
  • Confusion matrix with interactive pop-up and percentage toggle
  • Create Map — generate a GeoTIFF classification raster from any trained classifier
  • Export labels as JSON, GeoJSON, or ESRI Shapefile (pixel labels vectorized to polygons via d3-contour); download trained models
  • Label sharing — contribute to the Tessera global habitat directory (private) or share with other users (public)
  • Zarr embedding store — pulls from dl2.geotessera.org zarr where available, falls back to per-tile NPY downloads

Quick Orientation

TEE/
  public/js/                 8 ES modules loaded by viewer.html
  public/schemas/            Built-in classification schemas (UKHab, HOTW, EUNIS)
  public/viewer.html         Main viewer (6-panel layout)
  api/views/                 Django view modules (HTTP endpoints)
  api/urls.py                URL routing
  lib/                       Pure-function backend libraries (paths, viewport ops, tile rendering, pipeline)
  packages/tessera-eval/     Standalone ML library used by the tee-compute server
  process_viewport.py        Pipeline script (subprocess) — fetches tiles → pyramids → vectors
  scripts/deploy-compute.sh  Start tee-compute (use --local for Django + tee-compute on localhost)
  docs/                      This documentation

The frontend is vanilla JavaScript with no build step. Modules communicate through window.* properties bridged via Object.defineProperty. The backend is Django served by Waitress WSGI; ML evaluation runs on a separate Flask service (tee-compute) which Django proxies on /api/evaluation/*. A background pipeline downloads embedding tiles via the GeoTessera library (zarr-first, NPY fallback) and builds pyramids + vector data.