Airfree
Spatial AI · Machine Learning · Explainability

Spatial AI — intelligence embedded in every layer

Airfree's spatial AI platform fuses satellite imagery, ground-truth cadastral data, and decades of temporal records into models that governments and enterprises can audit, trust, and act on — at national scale.

Core capabilities

What our models do

Six production AI capabilities deployed across national land administration, environmental monitoring, and infrastructure management programmes worldwide.

Land Classification

Automated pixel-level classification of land cover and land use from multi-spectral and hyperspectral imagery — distinguishing agricultural, urban, forest, wetland, and bare-earth categories.

Change Detection

Temporal analysis across multi-date satellite archives to identify new construction, deforestation events, coastal erosion, and land-use transitions at national scale within hours of imagery ingestion.

Anomaly Detection

Statistical and deep-learning models surface statistically significant deviations from historical baselines — identifying illegal land clearing, informal settlement expansion, or infrastructure degradation automatically.

Object Detection in Satellite Imagery

Sub-metre resolution convolutional models detect and count buildings, vehicles, vessels, solar installations, and agricultural structures directly from optical and SAR satellite imagery.

NLP for Cadastral Documents

Large language models trained on government land records extract structured parcel data, ownership chains, encumbrances, and legal descriptions from unstructured historical deeds, survey plans, and title instruments.

Predictive Analytics

Time-series forecasting models trained on environmental, economic, and spatial covariates project land value trajectories, infrastructure demand, flood inundation extents, and agricultural yield outlooks.

Model architecture

The right model for every problem

Airfree maintains four families of production-grade models, each selected, trained, and validated for the specific statistical structure of geospatial problems.

Supervised Classification

Random Forest · XGBoost · CNN ensembles

Trained on labelled feature stacks combining spectral bands, elevation derivatives, and texture indices. Optimised for high-recall government workflows where false-negative costs exceed false-positive costs.

Land CoverSoil TypeVegetation ClassImperviousness

Semantic Segmentation

U-Net · SegFormer · DeepLab v3+

Fully convolutional models operating at pixel level over very-high-resolution imagery. Purpose-built for cadastral boundary delineation, building footprint extraction, and road network vectorisation.

Building FootprintsParcel BoundariesRoad NetworksWaterbodies

Time-Series Forecasting

Temporal Fusion Transformer · LSTM · Prophet

Multi-variate sequence models trained on dense temporal stacks of satellite observations, weather reanalysis data, and ground sensor readings. Delivers probabilistic forecasts with calibrated uncertainty intervals.

Crop YieldFlood RiskLand ValueVegetation NDVI

LLMs for Document Processing

Fine-tuned transformer models on cadastral corpora

Domain-adapted language models trained on large corpora of Australian and international land administration documents. Extract, normalise, and reconcile structured records from heterogeneous archival sources.

Title ExtractionDeed ParsingLegal Description NLPOwner Chain

Training data

Models trained on geometrically accurate, government-grade geospatial ground truth

Accurate spatial AI begins with ground truth that internet-scraped models lack. Airfree's training pipeline is built to assemble, label, and quality-control geospatial corpora sourced from national mapping agencies, land registries, and agricultural survey programmes.

Unlike foundation models trained on internet imagery, the platform's datasets are designed to be geometrically accurate, jurisdiction-stamped, temporally indexed, and enriched with government ground-truth — so models perform in production rather than just on benchmark leaderboards.

All training data is sovereign-cloud-resident, lineage-tracked, and governed by data-sharing agreements with the authorities that originate it.

Vector + raster

Labelled geospatial feature types

Polygons, points, and raster segments

30+

Years of temporal coverage

Multi-temporal Landsat, Sentinel, and commercial imagery

Government-grade

Ground-truth reference data

National survey and cadastral sources

Lineage-tracked

Training-data governance

Indexed, versioned, and sovereign-cloud-resident

Explainable AI

AI decisions that governments can stand behind

AI outputs that affect property rights, environmental enforcement, or infrastructure investment are legal instruments. Airfree designs every production model pipeline with auditability and defensibility as first-order requirements — not afterthoughts.

Our XAI framework meets the evidentiary standards of Australian administrative law, EU AI Act Annex IV documentation requirements, and Singapore's Model AI Governance Framework.

Confidence Scores on Every Prediction

No model output enters a government workflow without a calibrated confidence interval. Operators see exactly how certain the model is before acting on a recommendation.

Saliency Maps and Attribution

For every image classification or detection result, LIME and SHAP attribution maps show which pixels or features drove the decision — enabling human reviewers to verify model reasoning.

Full Audit Trails

Every AI-assisted decision is logged with model version, input hash, confidence score, and operator action. Audit logs are immutable, sovereign-cloud-resident, and exportable for regulatory review.

Model Cards and Datasheets

Every production model ships with a published model card documenting training data composition, known failure modes, demographic and geographic performance disaggregation, and intended use constraints.

Human-in-the-Loop Workflows

Below configurable confidence thresholds, predictions are automatically escalated to qualified human reviewers before producing legal or administrative effect — AI assists, humans decide.

Independent Accuracy Validation

All production models undergo third-party accuracy assessment against held-out ground-truth datasets prior to government deployment, with results published in our transparency reports.

Use cases

AI solving real government problems

Four deployment patterns that show how Airfree's spatial AI delivers measurable outcomes across land administration, disaster resilience, enforcement, and agriculture.

Automated Land Titling

AI-assisted extraction of parcel boundaries and ownership records from undigitised cadastral archives accelerates land formalisation programmes — reducing manual processing time by up to 85% in pilot deployments across three Pacific jurisdictions.

Flood Risk Prediction

Ensemble hydrological and ML models ingest DEM, soil moisture, rainfall forecast, and upstream gauge data to produce 72-hour flood inundation maps with property-level resolution — enabling emergency planners to pre-position resources before events.

Illegal Construction Detection

Change detection models run nightly over high-resolution imagery to identify new structures that lack corresponding permit records — automatically generating enforcement referrals to local authority systems via the Airfree Cadastre API.

Crop Health Monitoring

Multispectral vegetation indices combined with soil-adjusted reflectance models produce weekly crop health dashboards for national agricultural agencies, forecasting yield variation across irrigation districts months in advance.

Research partnerships

Academic foundations, production results

Airfree's AI capabilities are grounded in peer-reviewed research conducted in partnership with leading universities and institutes across Australia, Europe, and Asia. Published findings inform every model we ship.

Research enquiries
Australia

University of Melbourne — Spatial Informatics Lab

Probabilistic land cover mapping and uncertainty quantification in deep learning

Active research partner
Australia

UNSW Sydney — Remote Sensing & Geospatial Engineering

SAR-optical fusion for all-weather change detection and 3D reconstruction

Active research partner
Australia

Australian National University — Research School of Earth Sciences

Climate-coupled land-use change modelling and long-term ecosystem monitoring

Active research partner
Switzerland

ETH Zürich — EcoVision Lab

Large-scale biodiversity monitoring from satellite time series using self-supervised learning

Active research partner
Netherlands

Delft University of Technology — Geoscience & Remote Sensing

Interferometric SAR for subsidence monitoring and coastal land dynamics

Active research partner
Singapore

National University of Singapore — Urban Analytics Lab

AI-driven urban growth modelling and informal settlement characterisation

Active research partner

See AI in action

Ready to bring spatial AI into your land administration or environmental programme?

Our geospatial AI team works alongside government agencies and enterprise clients from proof-of-concept through to national-scale deployment — with explainable outputs your teams can defend and your stakeholders can trust.