Perm scientists have taught a neural network to search for illegal dumps in a new way—accuracy has increased from 56 to 95%.
Scientists Perm National Research Polytechnic University (PNRPU) have developed a new method for automatic analysis of satellite images.

Previously, neural networks analyzing satellite images often confused garbage deposits with construction sites, quarries, and industrial zones, forcing environmentalists to manually recheck thousands of images. Perm researchers changed their approach to training the model.
How Perm researchers trained a neural network
The neural network was trained on a base of more than 100 thousands Images from various regions of Russia were specifically included in the sample, including images of industrial and construction sites resembling landfills, so that the algorithm could clearly distinguish them from actual waste.
Furthermore, the scientists created a program that automatically collects additional terrain data from open sources, including topography, land cover type, and population density. This allows the neural network to evaluate not only the "picture" but also the characteristics of the surrounding area.
"During testing, we compared the standard training approach with the new one—using additional examples of objects that resembled landfills but weren't. As a result, recognition quality significantly improved. The rate of false positives—instances where the neural network mistakenly identified an area as a landfill—decreased by approximately 24 times. Overall classification accuracy increased from 56% to 95%. — said the author of the work, Vadim Danelyan.