TerraLuma · University of Tasmania
Introduction
Hello,
I'm James.
- Research Associate in Drone Data Platforms with the TerraLuma Group, University of Tasmania.
- Recently completed my Masters in Environmental Geospatial Science
- Previously worked as a software engineer of web applications & data visualisation.
Introduction
The NatureScan Project
- Innovation grant project
- Consumer-grade drones for remote sensing of biodiversity
- Aligned with the emerging biodiversity repair market
- Drone surveys completed across multiple sites around Australia, representing a range of different ecosystems
- One goal: an AI model that takes drone imagery and predicts Essential Biodiversity Variables (EBVs)
The Problem
We need labelled training data
To train or validate biodiversity models, we need datasets where individual life forms are both outlined and classified.
The Problem
Classifying
Segmenting
Potential Solution
Segment
Anything
Model
Potential Solution
The SAM Family
Introduced Promptable Visual Segmentation (PVS). Point, box, or multi-point prompts → mask.
Extends SAM to video, tracks and segments objects across frames. Also improves PVS performance.
Adds text-aligned concept understanding. Can respond to semantic descriptions. SAM3 still does and improves PVS (yes but no).
Geospatial Nuances
Image Size
- SAM is optimised for images around 1,000 × 1,000 pixels — it internally rescales to this range
- Fine for standard photos (500–4,000px), rescaling has little impact
- Problem: An orthomosaic is typically 15,000 × 15,000px or larger. Rescaling down destroys a lot of detail.
- Solution: Tile or clip into overlapping sub-images. Process each tile, then re-assemble in georeferenced space
Geospatial Nuances
Coordinate Systems
- SAM works entirely in pixel coordinates. All inputs (prompts) and outputs (masks) are in pixel space
- Our data is in a geographic CRS (e.g. GDA2020).
- On a webmap, our data is also projected into Web Mercator.
- Image pixels are flipped on the Y axis relative to geographic coordinates.
Geospatial Nuances
Multispectral Data
- SAM accepts only 3-channel, 8-bit input (like a JPEG or PNG)
- Our MS camera captures Green, Red, Red-Edge and Near-Infrared (NIR) bands as 32-bit floats.
- Workaround: false colour composites. Render and scale a 3-band combination into 8-bit RGB.
Results
How well does it work?
Tested on our orthomosaic data using manually segmented trees as ground truth.
SAM tends to under-segment, capturing the core of a plant. Conversely, I tended to oversegment in my manual process.
Demo
Resources
Next Steps
TerraLuma · University of Tasmania
Thank
you.
Slides QR