TSU will teach artificial intelligence to understand agriculture

An interdisciplinary team of TSU scientists - soil scientists, physicists, and meteorologists - is developing a system of markers for use in precision farming. The main source of information is Earth remote sensing data - satellite images, in which researchers single out visual characteristics that are special for particular soil indicators. The library of data collected will be the basis for training a computer model that will conduct digital analysis of the fields and help to increase the yield on them.

- The physical and chemical characteristics of the soil are the main parameters on which the condition of crops depends, the dynamics of their growth and ultimately the yield, and therefore the productivity and profit of agricultural companies, explains Oleg Merzlyakov, the project manager, associate professor in the TSU Department of Soil Science and Soil Ecology.

The approach developed by TSU scientists will help to get a holistic picture of the entire field, regardless of its scale. The new method is based on the analysis of spectral optical reflectivity and soil indices. The reflectivity coefficient provides information on the amount of humus in the soil, its particle size distribution, degree of moisture, and other agrophysical indicators. The information, with coordinates, helps to identify problem areas of the field, for example, where the soil has high acidity or insufficient nitrogen content.

At the stage of creating the marker library, active fieldwork will go on. Field visits are necessary for scientists to determine whether identification by satellite images corresponds to reality. When the data library is sufficiently populated, the design team will move on to machine learning of the computer model. Artificial intelligence based on the available material will learn to isolate specific visual characteristics that indicate the lack of certain elements and the state of the soil in each area. Since all sections are referenced to the coordinate system, the final result of the analysis carried out by the neural network will show the “weaknesses” with such a linkage.

- This image can be uploaded to unmanned agricultural machinery and used, for example, for targeted fertilizer application. It’s economically and environmentally more profitable than sprinkling fertilizer over the entire field, as is often done now. This not only does not give the desired result but can also adversely affect crops since an excess of trace elements is as undesirable as their shortage, says Oleg Merzlyakov. - Data obtained using the new approach may have another point of application, for example, to analyze plots from which less crop has been harvested. If you overlay a digital image obtained after analysis by AI, you can determine the causes of low yields and choose solutions to this problem.

Digital agricultural technologies are of particular importance for the Tomsk Region, because the region is a zone of risky agriculture. New approaches and tools will increase the yield by 20-25%, improve its quality, significantly reduce the cost of fuel and lubricants, reduce the environmental burden on the soil, and maintain its fertility.