Researchers from the Skoltech (part of the VEB.RF group) Artificial Intelligence Center analyzed more than 200 studies on the use of satellite data and artificial intelligence to monitor wildfires, floods, droughts, and land-cover change. They compared approaches used at different stages in the field’s development, from spectral indices and physical models to deep neural networks, transformers, and geospatial foundation models. The review was published in Earth Systems and Environment, a Q1 journal published by Springer Nature.
Satellite systems provide regular observations of vast territories and information that is difficult to collect on the ground. These data are much more than ordinary photographs of Earth. Optical, radar, thermal, and laser instruments capture different properties of the surface and atmosphere. Modern AI models can combine these observations and identify signs of fire, flooding, drought, deforestation, forest recovery, and other changes.
The authors traced how methods for processing such observations have evolved. Early studies mainly relied on spectral indices — formulas that compare measurements in different wavelength bands and help identify water, vegetation, or burned areas. Classical machine-learning methods followed, and then deep neural networks capable of learning relevant features directly from large datasets.
Different architectures address different parts of the problem. Convolutional networks capture the spatial structure of an image, recurrent models track processes over time, and transformers connect information across distant areas and long observation sequences. Geospatial foundation models are pretrained on large archives of satellite data and then adapted to specific tasks, such as flood mapping or post-fire assessment.
There is no universal algorithm for every environmental hazard. Wildfires and floods can develop within hours, while drought, forest degradation, and land-cover change emerge over months or years. Some models therefore need to recognize changes quickly in newly acquired images, others must analyze long time series, and the most advanced systems must account for both space and time.
Deep learning offers clear advantages: models can automatically extract features from multispectral imagery, account for the context around each location, track both rapid and slow processes, and integrate data from multiple satellites and sensors. Yet strong performance on a test dataset does not necessarily mean that an algorithm is ready for operational monitoring.
One of the review’s central conclusions is that a more complex model is not necessarily a more reliable one. An algorithm trained in one region may lose accuracy elsewhere because of differences in climate, terrain, surface types, satellite instruments, and data quality. Many systems also provide a prediction without indicating how much confidence users should place in it, and their outputs may conflict with known physical relationships.
Some limitations cannot be removed simply by scaling up a model. Fire can rapidly change direction because of wind and terrain, so even the most sophisticated system cannot always predict how a wildfire will develop. The result also depends on the accuracy of weather forecasts and the completeness of satellite observations. In other words, the limiting factor is sometimes not algorithmic power, but whether enough information is available for a reliable forecast.
The authors also point to a gap between research results and real-world warning systems. Satellite data are widely used to detect events and map areas that have already been affected, but they are incorporated into operational forecasting much less often. For practical deployment, models must work reliably with continuously arriving data, maintain performance in new regions, and report the uncertainty associated with each prediction.
“We wanted to do more than list the latest algorithms. Our goal was to understand where they offer a genuine advantage and where they run into limitations imposed by the data and by environmental processes themselves. Operational monitoring requires more than high test-set accuracy: a model must work in new regions, respect the physics of the process, and show how much confidence users should place in its prediction,” said Svetlana Illarionova, Head of the Computer Vision for Data Processing Research Group at the Skoltech AI Center.
The authors argue that the most reliable path is not to replace physical models with neural networks, but to combine their strengths. Such systems should incorporate physical laws, integrate data from different satellites and ground-based sensors, retain accuracy in new regions, and accompany each prediction with an indication of its reliability.
The new review provides a broader picture of which methods already work in environmental monitoring, where their limits lie, and what challenges remain. Read more in Skoltech’s materials on the key problems of geospatial modeling and the earlier research on wildfire occurrence prediction.
The Skoltech AI Center authors are Svetlana Illarionova, Head of the Computer Vision for Data Processing Research Group; Usman Tasuev, Software Engineer; Ivan Rubin, Research Engineer; Dmitrii Shadrin, Head of Direction; Aleksey Zaytsev, Associate Professor at Skoltech and Head of the Skoltech–Sberbank Applied Research Laboratory; Alexander Marusov, Head of Research Group; Skoltech Professor Alexander Bernstein; and Professor of the Russian Academy of Sciences Evgeny Burnaev, Skoltech Vice President for Artificial Intelligence and Director of the Skoltech AI Center.
The study also involved Polina Tregubova, an expert in biological sciences, and Dmitrii Katalevsky, Head of the Innoscale platform and Founder and CEO of Digital Vision Solutions. Svetlana Illarionova and Usman Tasuev contributed equally to the work.
The work was supported by the grant for research centers in the field of artificial intelligence provided by the Ministry of Economic Development of the Russian Federation under Agreement No. 000000C313925P4F0002 and Skoltech Agreement No. 139-10-2025-033.
