Yandex Cloud showed how CatBoost finds hogweed in satellite images

Yandex Cloud demonstrated how it automated the search for Sosnowsky's hogweed in satellite imagery using CatBoost. The project is particularly timely following stricter regulations: from March 1, 2026, plot owners are obligated to monitor the spread of this plant.

Why This Matters

Hogweed ceased being a local problem for dacha owners long ago. It rapidly captures large territories, displaces other plants, and poses risks to people, which is why it has now come under the attention of regulators. When dealing with thousands of hectares, walking the terrain on foot or manually marking up images is too expensive and slow.

Satellite images provide scale, but without automation they still require extensive manual work. This is precisely where Yandex Cloud, together with students from the School of Data Analysis and the volunteer movement "Stop Hogweed," assembled a practical ML case. The team sought not merely to recognize the plant in individual frames, but to build a reproducible technical process that could be transferred to other remote monitoring tasks.

This approach matters for municipalities, environmental initiatives, and landowners who need regular monitoring, not a one-time check.

How the Pipeline Works

At the core of the solution is a classical but well-designed computer vision pipeline for remote sensing data. First, images in GeoTIFF format are standardized, then features are extracted from them, after which the model learns to distinguish hogweed areas from the rest of the landscape. Special emphasis is placed on the fact that the system works not only with raw pixels, but also with features that help better capture the characteristic structure of vegetation.

• normalization and preparation of GeoTIFF files
• marking hogweed outbreak areas on satellite images
• computation of color and spectral features, including the CIVE index
• training a CatBoost model for area classification
• transferring the approach to searching for other objects, from forest clearcuts to ruins

"Your own data center won't be necessary—this can be done at home."

For practitioners, this is perhaps the most important part of the story. Yandex Cloud essentially demonstrates that such projects no longer require enormous infrastructure or expensive teams of narrow specialists in satellite data. If there is access to images and sufficiently high-quality markup, a working model can be assembled in a relatively compact environment. This lowers the barrier to entry for small research groups, eco-activists, and regional teams.

Why CatBoost Was Chosen

One of the most interesting conclusions from the material is that not every satellite image task automatically requires a neural network. In the hogweed case, gradient boosting on well-prepared features proved to be very competitive. For such scenarios this makes sense: data are often limited in volume, markup is expensive, and interpretability and speed of experiments matter as much as final quality.

CatBoost wins here due to a simpler training cycle and lower computational requirements. Equally important is the broader takeaway: the same stack can be applied not only to invasive plants. The article directly states that such an approach is suitable for detecting forest clearcuts, destroyed structures, and other objects visible in aerial or satellite imagery.

Essentially, this discusses a template for applied geospatial ML: collect a labeled dataset, select informative features, and train a model that then scales to large territories.

What This Means

Yandex Cloud presented not just an educational experiment with CatBoost, but a rather practical model for territory monitoring. Against the backdrop of new requirements for plot owners, such tools can quickly transition from the realm of research projects to regular operational processes for business, regions, and environmental services.