Translation. Region: Russian Federation –
Source: Peter the Great St. Petersburg Polytechnic University –
An important disclaimer is at the bottom of this article.
Russian scientists have developed a method for detecting agricultural plant diseases at an early, asymptomatic stage. The approach is based on the analysis of hyperspectral data using artificial intelligence. The proposed approach has demonstrated the feasibility of detecting wheat stem rust, one of the most dangerous crop diseases affecting plant stems and leaves. This technology paves the way for the development of satellite and drone monitoring systems for preventive crop protection. The results of the study were published in the international scientific journal Frontiers in Plant Science. The research was supported by the Russian Science Foundation.
Wheat is one of the world's key grain crops, producing over 770 million tons of grain annually. Many varieties are susceptible to stem rust, which can cause significant yield losses. The situation is further complicated by the rapid interregional spread of aggressive pathogens. The challenge for agricultural sustainability lies in the fact that the effectiveness of protective measures is largely determined by the ability to detect infection before visually discernible symptoms appear. Typically, visually discernible symptoms appear only 6–10 days after infection. The objective of this study was to develop approaches for the early detection of plant diseases, enabling rapid localization of the infestation and minimizing losses for agricultural producers.
Scientists from the Advanced Engineering School "Digital Engineering" at Peter the Great St. Petersburg Polytechnic University and the All-Russian Institute of Plant Protection have proposed a method for the early detection of agricultural plant diseases. It is based on the use of artificial intelligence to process hyperspectral imaging data. This technology records light reflection in tens and hundreds of narrow spectral bands, enabling the detection of early physiological changes in plants even before the disease becomes visible.
The researchers conducted experiments on wheat plants grown under laboratory conditions similar to field conditions. The experimental datasets were acquired using a hyperspectral camera. A total of 864 hyperspectral images were collected, including both healthy and infected plants.
The main drawback of currently available remote sensing methods for plant assessment is that the resulting images do not always provide researchers with comprehensive data for analysis. Therefore, scientists from St. Petersburg have developed a method that relies primarily on the controlled collection and processing of primary visual information in real, challenging agricultural crop conditions, regardless of external factors.
When developing the new methodology, we took into account key challenges of plant remote sensing encountered in real agricultural conditions, including uneven lighting, overlapping vegetation structures, environmental humidity, background noise, and daily variability in data acquisition conditions, commented Anton Terentyev, a researcher at the All-Russian Institute of Plant Protection.
A key element of the developed methodology was the creation of an algorithm for sequential preprocessing of hyperspectral data that is robust to distortions arising during the acquisition process. Using artificial intelligence and machine learning tools, an algorithm with formalized stages, relationships, and reproducible procedures was developed. The most important quality criteria for the algorithm were the reliability of the experimental hyperspectral data processing results and high processing speed. The published scientific article demonstrates that properly organized data preprocessing plays a key role in improving classification quality and the stability of results, regardless of the model used.
The key factor in the method's effectiveness was not the model complexity, but rather the correct data preprocessing, which allows machine learning algorithms to reliably distinguish between healthy and diseased plants under various noise conditions. "We deliberately emphasized the interpretability of the AI models' decisions, since without understanding the basis on which these models make decisions, the risk of errors increases," emphasized Alexander Fedotov, leading researcher at the Advanced Engineering School's "Digital Engineering" laboratory.
The authors note that the developed method can be practically implemented in remote monitoring systems for agricultural land, including unmanned and satellite platforms, for the early detection of other diseases and stress conditions in agricultural plants.
The study was supported by the Russian Science Foundation (grant no. 25-21-00444).
Please note: This information is raw content obtained directly from the source. It represents an accurate account of the source's assertions and does not necessarily reflect the position of MIL-OSI or its clients.