MTS will train AI to combat Sosnovsky hogweed

As TAdviser found out , MTS decided to create a dataset for training a neural network to recognize Sosnovsky's hogweed. From the documentation for the purchase of services for the collection of such a datacet, posted in mid-February 2026^[1]it follows that it will be designed to train and validate the computer vision model, capable of detecting Sosnovsky's hogweed with an accuracy of 95% or higher in images under various conditions.

The terms of reference specify that the areas of application of the datacet are: automated monitoring of the contamination of territories, the formation of maps of the spread of an invasive species and support for solutions for phytosanitary measures.

The data sources for the data set during the initial collection will be field surveys covering the regions in the Central Federal District, the Northwestern Federal District, the Volga Federal District, the Southern Federal District, and shooting from drones equipped with RGB cameras. Photographs from UAVs from different heights are required - from 50 to 150 meters. Data from open sources will also be used, such as crowdsourcing platforms (iNaturalist, GBIF) - with verification of species identification, and archives of regional phytosanitary services.

𝓲

Фото: Shutterstock

We consider all areas that are complementary to our business, including solutions related to environmental risk management. The hogweed recognition project is one of them. At the moment, we are only studying its prospects, "TAdviser explained to MTS the prerequisites for this project.

For MTS, this is not the first experience in terms of AI solutions related to environmental risk management. Earlier in its portfolio, for example, a system of proactive environmental risk management has already appeared, which is able to analyze the environmental situation in a short time, make dynamic forecasts of the state of the air basin and identify the contribution of specific sources to atmospheric pollution. According to MTS, as of mid-2025, this solution was introduced in 14 regions of Russia in the field of metallurgy, fertilizers, pulp and paper industry and in others^[2] of^[3] Trade].

The spread of Sosnovsky hogweed is a long-standing and well-known problem that poses a threat to agriculture. To combat it, various measures are applied, including legislative and technological. In some regions, AI has already begun to be used. So, for example, in 2025, a pilot project was conducted in Tula, in which AI using satellite data automatically identifies the foci of a dangerous plant, conducts a survey and proposes a plan to combat hogweed.

In the same year, AI began to be used in St. Petersburg to identify Sosnovsky hogweed. Moreover, in a fairly short time, the first few owners of land plots received fines for maintaining the territory due to hogweed, which was discovered using the neural network^[4].

The definition of hogweed by a neural network is a rational and practical task, but it will give the maximum return not as a "separate application," but as a service in the agricultural processes already used by agricultural scouting: monitoring of land, staging work, execution control, re-verification, believes Igor Shupenev, deputy general director of Agropromcipra. The key to value is the delivery of the result to each agricultural producer: an understandable map of foci, prioritization by risks and immediately tasks for processing with confirmation.

It is important to consider adjacent sections and propagation corridors (field boundaries, roads, ditches, watercourses) to see the risk of re-skidding. And the economy needs to be embedded in current processes, software and budgets, and not made a separate line of costs, otherwise the main risk of scaling is not in the accuracy of the model, but in the delivery of the solution and an understandable ROI for the economy, - said Igor Shupenev in a conversation with TAdviser.

Notes