Researchers at Perm National Research Polytechnic University (PNIPU) developed a new data preparation method for neural networks to improve the detection of illegal waste dumps. By optimizing the training process, the team increased the classification accuracy from 56% to 95%.
The researchers utilized over 100,000 satellite images from various Russian regions for training. To improve performance, they created a program that automatically gathers geospatial data, such as terrain type and population density, from open sources to provide context for the model.
The team trained a standard image analysis neural network on an expanded dataset that included satellite images of quarries, industrial zones, and construction sites. This approach helped the model distinguish between actual waste sites and visually similar objects.
According to the study, the new method reduced the rate of false positives by approximately 24 times. The findings were published in the journal Applied Mathematics and Control Issues, highlighting the potential for environmental monitoring services to streamline site inspections.
