v1.2.1

v1.2.1

Label Export/Import

• Vectorized export: pixel labels exported as polygons via d3-contour marching squares
• Shapefile import: JSZip extraction, point grouping by name, polygon rasterization fix
• Panel 6 delete fix: last label deletion now updates both panels

User Management

• Email field in user enrolment (UI, API, and CLI)
• Zulip welcome on first sign-in
• Restored Users tab UI (lost during earlier refactoring)

User Guide

• Complete rewrite for non-technical ecology audience
• Screenshots, Mermaid diagrams, TOC, numbered sections, PDF version
• Classification schemas, classifier parameters, spatial split documented

Labelling

• Store cluster pixel_coords on promote
• Multiple shapefile uploads with Clear button
• EUNIS habitat classification schema

Infrastructure

• Unified deploy-compute.sh
• Zarr support with chunked fetching
• 19 new validation tests

v1.2.0-beta

Beta 1.2.0

Label Export/Import

• Vectorized export: pixel labels are now exported as polygons (via d3-contour marching squares) instead of individual points — reduces 216K features to a handful of polygons per class
• Shapefile import fix: uses JSZip to extract .shp/.dbf from zip (was passing raw zip bytes)
• Import grouping: multiple points with same name are reconstituted as a single label with pixel_coords, centroid embedding, and threshold
• Panel 6 last-delete fix: deleting the last label now correctly clears the panel 6 mirror

Labelling

• Store cluster pixel_coords on promote for exact rebuild on show/hide
• Clear localStorage labels when viewport is deleted
• Fix panel 6 header hiding the auto/manual label dropdown
• Allow multiple shapefile uploads with explicit Clear button

Evaluation

• Create Map: GeoTIFF generation from classifier predictions
• EUNIS terrestrial habitat classification schema (6 groups, 35 level-2, 246 level-3)
• Zarr support with chunked fetching and coverage probing
• Shared zarr utilities (get_zarr, probe_zarr_coverage, read_region_chunked)

Infrastructure

• Unified deploy-compute.sh with --local flag (replaces restart.sh)
• Port 8001 for tunnel — no Django needed for evaluation
• Zarr-first viewport processing with NPY fallback
• 19 new validation tests for export/import pipeline

Beta 1.1.0 — Spatial evaluation, hyperparameter sweep

What's New

Spatial train/test split

• Draw train (blue), test (yellow), and map (green) bounding boxes on the satellite map
• Training points sampled from train areas only; test set is fixed from test areas
• Avoids spatial autocorrelation for more realistic accuracy estimates
• Bounding boxes saved in config files
• Warning when test set is too small (<100 pixels)

Hyperparameter sweep

• Click "+" next to any classifier's parameters to add variants
• Each variant gets its own learning curve (e.g., "MLP (v1)", "MLP (v2)")
• Config file supports list format: "mlp": [{"hidden_layers": "64,32"}, {"hidden_layers": "256,128,64"}]
• Variant-aware chart colors (lighter/darker shades) and labels

Performance improvements

• Vectorized point extraction from tiles (~100x faster per tile)
• GDF filtered by tile bbox before CRS reprojection (~40x faster)
• Max pixel samples now controls actual sampling budget

UI improvements

• Crosshair cursor when drawing rectangles
• Auto-draw: selecting area type starts drawing immediately
• Map dragging disabled while drawing (no pan/draw conflict)
• Config file includes inline documentation of choices

Bug fixes

• Spatial MLP IndexError: separate data pool for spatial classifiers
• UnboundLocalError for spatial_labels on cache hit
• Deploy script: kill -9 for reliable process termination

Deployment

# Hosted server
docker pull sk818/tee:stable
sudo ./manage.sh   # option 7

# Compute server
./scripts/deploy-compute.sh gpu-box

Beta 1.0.0 — Large-area evaluation with GPU compute

What's New

Large-area evaluation

• Evaluate classifiers on country-scale shapefiles (tested on Austria: 42K features, 40 tiles)
• GeoTessera point sampling for pixel classifiers (k-NN, RF, XGBoost, MLP)
• Real tile-aligned 256×256 patches for U-Net and spatial MLP
• Configurable sampling strategies: equal, sqrt-proportional, proportional to area

Remote GPU compute

• tee-compute server for offloading ML to GPU boxes via SSH tunnel
• Deploy script for easy server management
• OpenBLAS fix for servers with >128 CPU cores
• Server-side cancellation (cancel button stops backend computation)

U-Net improvements

• Pixel-aligned patches from real GeoTessera tiles (not point grids)
• 16× augmentation (4 rotations × 2 flips × 2 noise levels)
• GPU support via PyTorch with CUDA

UI improvements

• Unified learning curve x-axis: % of labelled pixels used for training (log scale)
• Scrolling progress log in validation panel
• Status button showing backend + compute hostnames
• Per-class polygon counts from full shapefile (not truncated GeoJSON)
• Sampling strategy dropdown (equal / sqrt-proportional / proportional)
• Max pixel samples and max patches controls

Data & caching

• Disk result cache: instant re-runs with same shapefile/field/year
• Single tile pass: point samples and patches extracted from same tiles
• Estimated total labelled pixels from polygon areas at upload

Example data

• austria.zip included — Austrian INVEKOS crop field data (42,789 polygons, 17 crop classes)

Deployment

# Hosted server (tee.cl.cam.ac.uk)
docker pull sk818/tee:stable
sudo ./manage.sh  # option 5

# GPU compute server
pip install -e packages/tessera-eval[server]
tee-compute --port 5050

See the User Guide for full setup instructions.

Alpha 3.7.0

What's New

Label Sharing

• Private mode: share embedding+label pairs with the Tessera team (no locations)
• Public mode: share geolocated labels as ESRI Shapefile for other users to import
• Import button shows available shared labels with badge counter

New Built-in Schema

• HOTW (Habitats of the World) added alongside UKHab

Architecture Cleanup

• Removed GeoTIFF tile codepath — PNG-only pyramids
• Removed legacy label system (~600 lines of dead code)
• Removed Stage 2 satellite pipeline
• Renamed maps.heatmap → maps.panel5, HEATMAP_LAYER_RULES → PANEL5_LAYER_RULES
• Enforced embeddings (uint8) vs vectors (float32) naming convention
• Simplified progress system to single source of truth
• Upgraded shp-write to v0.4.3

Bug Fixes

• Polygon drawing: fixed cross-module bare calls, latlng unwrapping, completion dblclick
• Map sync: fixed missed panel5 renames in sync arrays
• Viewport re-creation: thorough cleanup of stale files, race conditions
• IndexedDB cache migration for .embeddings → .values rename
• Shapefile upload: read all .shp files in ZIP (points + polygons)
• OSM attribution compliance
• Safari fixes: Float32Array JSON serialization, shp-write MIME type

Documentation

• Comprehensive docs: architecture, frontend API, backend API, extension guide
• Schema file format documentation with examples
• Known technical debt documented (dual label systems, double quantization)

Viewport Creation

• Removed blocking GeoTessera registry check (was 28s delay)
• Granular progress: shows GeoTessera init time, download MB/speed
• GeoTessera availability check deferred to pipeline

Other

• Smaller polygon vertex markers (1/4 size)
• Panel titles: "Create labelled points", "Classification results"
• tile_health auto-discovers year directories

v3.5.0 — Django Auth Migration

What's changed

Replaces the custom /data/passwd + bcrypt authentication with Django's built-in contrib.auth backed by SQLite, following Anil Madhavapeddy's security review.

Highlights

• Django admin UI at /admin/ for managing users and quotas
• Automatic passwd migration — existing /data/passwd files are imported on first container boot (no manual steps)
• Management commands: tee_adduser, tee_removeuser, tee_listusers, tee_setquota, migrate_passwd
• PBKDF2 password hashing (default) with BCrypt support for migrated hashes
• Session fixation protection via session.cycle_key() on login
• No frontend changes — API response shapes are identical

Upgrade

On the host VM:

sudo ./manage.sh   # option 5 (Update container)

The container auto-runs migrate, collectstatic, and migrate_passwd --auto on startup.

Improved error messages

• process_viewport.py now shows friendlier retry/failure messages

v3.4.0

v3.4.0

Manual Labelling Mode

• Pin labels — Ctrl+click (or double-click) to place point labels with embedding capture
• Polygon labels — Ctrl+double-click to draw polygon boundaries, rasterized at 10m resolution
• Per-class similarity sliders — expand label coverage by adjusting embedding distance thresholds
• Classification — click Classify for full-viewport nearest-centroid classification in Panel 5
• Panel 4 integration — manual label classes colored in PCA/UMAP scatter plot
• Timeline — track label coverage across years with clock icon
• Schema support — UKHab v2 or custom JSON/CSV for standardized class names

Auto-label Improvements

• Rename promoted labels with pencil icon
• Eye icon replaces Hide/Show text, clock icon replaces Timeline text
• Fix: promoted cluster pixels no longer inflate after viewport switch

Validation

• Export Results and Download Models buttons moved to top header bar
• Spatial MLP (3×3 and 5×5) and U-Net (GPU) classifiers added

User Guide

• Comprehensive rewrite: documents all 4 viewer modes (Explore, Change Detection, Labelling, Validation) with per-panel descriptions
• Manual labelling guide: pins, polygons, thresholds, classification, Panel 4 coloring
• Export docs split for header vs manual label exports (JSON, GeoJSON, ESRI Shapefile ZIP)

Bug Fixes

• Fix Panel 4 colors disappearing on viewport return
• Restore threshold-gated classification (prevents marking all pixels)
• Set-active icon moved to leftmost position in label rows

v3.3.0 — Labelling UX improvements

What's new

• Polygon drawing fix — vertex clicks no longer create spurious point rows in the labels list
• Smooth similarity slider — debounced with requestAnimationFrame; no DOM rebuild mid-drag
• Label row layout — pixel count moved to right of slider; unified layout for polygon and point rows
• Scrollbar fix — labels list scrollbar no longer overlaps the delete button
• Toolbar cleanup — removed redundant "View Labels" and "Save as Label" buttons
• Satellite acquisition date — ESRI imagery date shown in Panel 2 header

v3.2.0

What's New

U-Net Classifier

• U-Net (GPU) classifier operating on the full 2D embedding grid (1, 128, H, W) via PyTorch
• Supports CUDA, MPS (Apple Silicon), and CPU
• Configurable epochs, learning rate, depth, and base filters
• 1 repeat per training size (vs 5 for sklearn classifiers)
• Trained models download as .pt files
• Greyed out in deployed mode (separate tile server)
• PyTorch imported lazily — not bundled in Docker image

Spatial MLP Improvements

• 4× flip augmentation for Spatial MLP (3×3) and (5×5) — horizontal, vertical, and both flips are label-preserving for land cover patches

Other

• TEE mosaic logo added to README
• Graceful handling of non-JSON server errors in the client

Full changelog: v3.1.0...v3.2.0

v3.1.0

What's New

Validation & Evaluation

• Spatial MLP (3×3) classifier with cancel button and sorted class table
• Spatial MLP (5×5) classifier with memory-efficient spatial feature gathering
• Streaming learning curves via NDJSON with per-classifier early finish
• Confusion matrix panel with metric toggle and CSV export
• Trained model download after validation runs
• Tunable hyperparameters for all classifiers

UI & UX

• Suggested labelling workflow in user guide
• Research disclaimer and acceptance checkbox on login
• Fixed mode-dependent panel visibility across all four viewer modes
• Fixed UMAP recomputing on every dropdown selection
• Fixed similarity search pins not clearing on subsequent double-clicks
• Fixed View Labels button not showing panel
• Removed 2017 from year selection (poor part-year Sentinel data)

Stability

• Restored fetch timeout and improved evaluation error messages

Full changelog: v3.0.0...v3.1.0