TerraLabel
An AI-powered geospatial labeling tool that transforms single clicks into accurate GeoJSON polygons using Meta's SAM2. Human-in-the-loop annotation at 5-10x the speed of manual digitization.
Platform
terralabel.ai
Platform Capabilities
Faster Than Manual
GPU Inference
Undo/Redo Steps
To Polygon
The Problem: Scaling Geospatial Annotation
Training ML models for remote sensing requires labeled data. Whether detecting solar panels, extracting building footprints, or mapping agricultural parcels, someone needs to trace those boundaries. Traditional approaches are either accurate but slow (manual) or fast but often wrong (fully automated).
Manual Digitization
Accurate but tedious. Hours spent tracing polygon boundaries pixel by pixel.
Fully Automated
Fast but unreliable. Models make mistakes that require manual correction anyway.
TerraLabel Approach
AI handles boundary tracing; humans handle judgment calls. Best of both worlds.
SAM2 Integration
TerraLabel leverages Meta's Segment Anything Model 2 (SAM2) with Hiera Large backbone. SAM2's image encoding is expensive, but subsequent prompts are cheap—perfect for interactive labeling workflows.
Click-to-Polygon
Single clicks are transformed into accurate polygons. The AI traces object boundaries automatically, understanding context from the surrounding imagery.
Hiera Large Backbone
SAM2's hierarchical vision transformer provides multi-scale features critical for accurately segmenting objects at various sizes in satellite imagery.
Interactive Labeling Tools
Built on deck.gl's editable layers, TerraLabel provides professional GIS-quality drawing tools in the browser. Switch between modes depending on the annotation task.
AI-Assisted Selection
Single-click SAM2 segmentation that traces object boundaries automatically with 100-500ms latency.
Manual Drawing Tools
Point, Line, Polygon, Circle, and Drag Draw modes for precise manual annotation when needed.
Undo/Redo History
50-step history with full state preservation. Never lose work due to accidental edits.
Measurement Tools
Distance, Area, and Angle calculations with real-time feedback as you annotate.
Multiple Selection Modes
Rectangle and Polygon selection for bulk operations on multiple features.
Vector Tile Visualization
View thousands of labeled features across entire study areas using Martin tile server.
Three-Tier Architecture
TerraLabel connects a Svelte frontend with deck.gl mapping, a FastAPI backend running SAM2 inference, and PostgreSQL with PostGIS for spatial data storage.
Frontend
Svelte 5 with deck.gl editable layers. Handles all map interaction and polygon editing.
Backend
FastAPI with SAM2 inference. Simple REST API: send image and coords, receive polygon.
Database
PostgreSQL with PostGIS stores imagery and labeled features with spatial indexing.
Tile Server
Martin serves vector tiles directly from PostGIS for efficient visualization.
Vector Tile Visualization
View all labeled features across the entire study area. Individual tiles are ~50KB compared to loading the full dataset at 160MB+.
Technology Stack
Key Learnings
Technical Insights
- SAM2 generalizes well to satellite imagery despite natural image training
- Sub-second latency is critical for maintaining labeling flow state
- Douglas-Peucker polygon simplification is essential for usable output
- Vector tiles transformed visualization from unusable to smooth
Production Results
- Thousands of solar panels labeled using ACT aerial imagery
- 5-10x faster than manual digitization with comparable accuracy
- Human-in-the-loop workflow produces better results than either approach alone
- GPU inference (100-500ms) vs CPU (2-5s) makes interactive use viable
Try the live demo
Click on satellite imagery to generate AI-powered polygons. Explore thousands of labeled solar panels across Canberra.