SFU: Method for preprocessing hyperspectral images

2025: Introduction of Hyperspectral Image Preprocessing Method

Scientists from SFU have developed an approach to working with hyperspectral images - images that allow you to "see" the chemical composition of objects based on their interaction with light. Hyperspectral images are used for field health monitoring, environmental research and food quality control. Such images may show, for example, the content of moisture or certain substances in the leaves of plants. The proposed approach, based on random selection of key spectral data from hyperspectral images, helps artificial intelligence analyze data more accurately, and also reduces the likelihood of errors by 15%. This was announced on June 18, 2025 by the Russian Scientific Foundation (RNF).

𝓲

Фото: Boris Kozlovsky

Hyperspectral images are photographs that show not only colors, but also spectral characteristics of objects, such as the presence of water, chlorophyll, the level of organic and mineral substances. Thus, objects with different chemical compositions reflect light in different ways, and therefore their hyperspectral images turn out to be individual.

Simply put, such images help to see what is invisible to the human eye, such as differences in the chemical composition of soil or leaves of plants. However, spectral data is too voluminous, so computers often do not cope with their processing: they can miss important signals, for example, the presence of nutrients in the soil or signs of stress in plants, or, conversely, find duplicate and irrelevant information. This makes it difficult to use hyperspectral data in real-world tasks, for example, to study the state of forests and natural ecosystems and detect pollution, oil leaks and other toxic substances in the soil.

𝓲

Фото: Boris Kozlovsky

In such a method, "bad" spectral profiles representing noise are mixed with "excellent" spectral profiles reflecting the true spectral state of the object. As a result, the researchers get "good" spectral profiles, which turn out to be closest to the arithmetic mean, modal or median spectral profile.

The use of spectral information pretreated in this way in machine and deep machine learning algorithms makes it possible to significantly increase the accuracy of assessing the state of plants. This approach can be compared to watching a photo album: instead of studying each photo, you can, using a statistical approach, choose only those that contain key points.

The proposed method helps to reduce the likelihood of errors. Thus, in experiments when using it, the accuracy of data analysis increased by 15% compared to other spectral methods of data processing. The advantage of the new approach is explained by the fact that the Random Reflection method reduces the complexity of hyperspectral information with almost no loss of data used for analysis.

Thanks to this approach, farmers will be able to more accurately determine which areas of the field require fertilizer or watering, which will lead to more efficient use of resources and increased yields. Environmentalists will be able to detect pollution in water and soil, and product manufacturers will be able to monitor the quality of raw materials and finished products, which can increase safety standards and consumer satisfaction. In addition, in geology, the method can be used to detect minerals and ores and map the area in order to determine the chemical composition of the soil, and in medicine - to diagnose skin diseases and study biological materials.

We plan to further test the Random Reflection method using various machine learning algorithms and solving various practical problems - to assess plant stress, classify the phenological states of conifers, identify species occupying new territories that are not characteristic of them, determine the moisture content of sunflower seeds and other crops. Hyperspectral technologies have huge potential, and our development contributes to their development, "said Boris Kozlovsky, candidate of biological sciences, senior researcher at the Academy of Biology and Biotechnology of the Southern Federal University, project manager supported by a grant from the Russian National Research Foundation.