# AI Tool for Automated Detection of Sosnowsky's Hogweed in Satellite Images
Developers from the Yandex School of Data Analysis, Yandex Center for Technologies for Society, and the StopHogweed movement have launched a computer vision-based service for detecting Sosnowsky's hogweed infestations. The tool speeds up image annotation 50 times compared to manual methods. It's free for environmental organizations, national parks, scientists, and farmers. The model was trained in Yandex Cloud on a dataset of 10,000 satellite images showing infestation hotspots. Validation used the IoU metric to assess detection accuracy, followed by expert review from StopHogweed specialists.
Technological Foundation and Testing
The service leverages proprietary Yandex computer vision models. The algorithm segments hogweed zones in high-resolution images. Volunteers tested the tool on data from 17 regions in European Russia. It detected 421 hectares of infestation, including full coverage of Moscow and Moscow Oblast. Based on the annotations, eradication efforts are underway: digging up roots and pruning inflorescences, mainly in protected areas.
A practical example is the detection of a major hotspot in Pleshcheyevo Lake National Park (Yaroslavl Oblast), which was successfully eradicated.
Using the Service in Practice
Users upload satellite images in GeoTIFF format with georeferencing. The algorithm returns annotations for infestation zones. Processed data is available on an interactive map. For unprocessed areas, the service generates segmentation automatically.
- Input Data Format: GeoTIFF with georeferencing metadata.
- Output: Vector contours of hotspots on the map.
- Advantages: Reduces annotation time by up to 50 times, scalable to large areas.
- Target Audience: GIS specialists, ecologists, volunteers skilled in satellite data.
Scaling Plans
In 2026, the service will expand coverage to 100,000 km² in Tver and Yaroslavl Oblasts. Developers will integrate data from Roszapovedtsentr for federally significant protected areas. Plans include training the model on other invasive species:
- Echinocystis lobata.
- Two species of goldenrods.
The goal is a single national map of hogweed hotspots. The tool maintains accuracy on complex landscapes, which is critical for monitoring invasive plants in GIS systems.
Key Points
- Speeds up annotation 50-fold thanks to an IoU-validated computer vision model.
- Detected 421 ha of infestation in 17 regions, full coverage of Moscow and Moscow Oblast.
- Scaling to 100,000 km² of protected areas by summer 2026.
- Expansion to echinocystis and goldenrods for comprehensive monitoring.
- Free access with GeoTIFF support for GIS integration.
The service demonstrates the power of ML in ecology, minimizing manual labor while upholding quality metrics.
— Editorial Team
No comments yet.