How we bring AI to the alpine pasture with Virtual Shepherd

“Bringing Innovation into Practice” was the motto of the AAIC – Applied AI Conference on 24 September 2026 in Vienna. In the joint session with Josephinum Research and Agro Innovation Lab, hosted by the Austrian Federal Economic Chamber (WKÖ) and ADVANTAGE AUSTRIA, our COO Marinca Kasiske used Virtual Shepherd to show how an AI project makes its way from research into everyday use.

Three hurdles we solve in the field

Anyone bringing AI to the field and the alpine pasture runs into three hurdles that don't matter in the lab: scarce data, missing connectivity & devices with limited power. Virtual Shepherd, a GeoAI project funded through the FFG ASAP programme, tackles the daily task of monitoring grazing livestock in alpine terrain. It is developed together with the University of Salzburg, TU Wien and GISolutions.

How we close data gaps with generative AI

Data scarcity is the rule in agriculture. Animal data is patchy and often unavailable at night. A generative AI, a Wasserstein GAN, fills these gaps with synthetic movement tracks and keeps the data base complete. Four-hour windows turn GPS points into heatmaps and dwell hotspots that reveal how the animals behave.

How we detect and explain anomalies

On this basis, an ensemble of three analytical AI models detects deviations in movement patterns. Across the two alpine field tests it detected 71 anomaly events. Each one is explainable via SHAP: the system shows not only that something deviates, but which factors contribute to it. From our point of view, this is the step from a simple alert to a decision you can trace.

How we handle connectivity & power

There is no reliable connection and no power socket on the pasture. Positioning therefore fuses GNSS, LoRaWAN & 4G/5G through a Kalman filter and reaches over 95 % accuracy in the Mölltal test area. Transmission runs over LoRaWAN and 4G/5G, with Starlink as a fallback in remote locations. And because power is scarce, part of the models learn directly on the device: TU Wien's federated learning reaches around 90 % of a central model's performance at over 50 % less energy.

What we contribute as the operator of the value chain

We see ourselves as the operator of the entire AI value chain. At MovingLayers, the threads come together: the AI platform with Azure cloud, Kubernetes, FROST as the IoT broker & PostGIS, the data pipeline with live data, heatmaps, virtual fences and AI-driven alerts, and connectivity from the pasture to the app. The data travels from the collar through the base station and the FROST backend into the VirtualShepherd app.

Where Virtual Shepherd fits into the geodata lifecycle

Virtual Shepherd does not stand on its own. In the geodata lifecycle it is one of two forms of the Capture phase: the data stream that runs without a crew and without a planned end, where mobile coverage cannot reach. The planned site visit with MoversSurvey is the other. The two forms belong together, because the data stream is often what shows that a site visit is needed: a sensor reports a deviation, the deviation triggers the site visit, and the site visit produces the reliable record.

This turns the SHAP-explainable anomaly detection into the trigger that starts the next cycle. What Virtual Shepherd captures on the pasture follows the same rules as every other state in the cycle: operated locally, traceable and kept so that no state is overwritten.

What we measured in the field test

The results so far come from two alpine field tests on the Zeiselalm (Lower Austria) and the Lorenzalm (Carinthia) (as of: AAIC presentation, 24 September 2026):

• 80,000+ location records from the two field tests, continuously hardening the models.
• 71 anomaly events detected, each explainable via SHAP.
• 65 % less cloud load through edge processing, with a geofence alert in under 500 ms.
• around 90 % model performance of federated learning compared to a central model, at over 50 % less processing energy.
• 3–4 hours of daily search time that farmers spend today, which the app can markedly reduce according to the field test.

What we are working on next

Virtual Shepherd is tested and scales on alpine pastures in Lower Austria and Carinthia. The next step is a LoRaWAN sensor array that extends the data stream of the Capture phase beyond positioning alone and bundles a wide range of information streams into one solution. This is how we address the data scarcity that stands in agriculture's way today, at its root.

Would you like to evaluate Virtual Shepherd for your operation? You'll find all the details on the Virtual Shepherd product page and how it fits into the wider process on the geodata lifecycle page. To arrange an initial call, get in touch here.

FFG project info: projekte.ffg.at