From satellite embeddings to Darwin Core: an open, ensemble-AI workflow for Prosopis juliflora mapping and reusable training-dataset generation

What this was about

Pranav Jha described mapping the invasive Prosopis juliflora (Neltuma juliflora) across India, especially about 1.1 million hectares in Kutch, Gujarat, using satellite data and geospatial foundation model embeddings. The workflow combines Sentinel, hyperspectral and NISAR radar data with ground truthing, AlphaEarth embeddings, NDVI and slope checks, reaching about 91% accuracy, and uses 3 m hyperspectral signatures to distinguish Prosopis from similar natives. Ground teams validate predicted clusters through drop, curation and refined zones to build reusable training datasets, and the model is being extended to Lantana camara and mangroves.

Why it matters. Links remote sensing AI to biodiversity data reuse, with a focus on transparent, reusable evidence for invasive species management.

In the room

• Prosopis juliflora is among IUCN's top invasive species and has spread across APAC and sub-Saharan Africa; in Kutch it covers about 1.1 million hectares, harming native species and pastoral communities, and has spread to 20+ agroclimatic zones.
• Satellite mapping is needed to understand invasion at scale, over time (multi-date imagery), where to look next and what it displaces.
• Density is mapped at 10 m resolution to guide ground teams to hotspots.
• Inputs: Sentinel, hyperspectral (e.g. AVIRIS and others), NISAR SAR data and ground truth; AlphaEarth geospatial foundation model embeddings train a classifier, followed by NDVI and slope/elevation checks.
• Current accuracy about 91%; working with Google DeepMind to include it in Google Earth.
• Prosopis resembles native Prosopis cineraria and acacias; 3 m hyperspectral spectroscopy separates them.
• Ground workflow of drop, curating and refined zones produces reusable datasets; goal is transparent, attributable, reusable evidence.

Notable moments