A team of Skoltech researchers has been granted a patent for a method of using neural networks to process remote sensing satellite images. This method, which uses computer vision to segment satellite images, will help make analytics based on remote sensing data faster and more accurate.

Skoltech Associate Professor Evgeny Burnaev, research scientist Alexey Artemov, research engineer Alexey Bokhovkin and PhD student Denis Volkhonskiy developed a new algorithm to train a neural network to recognize boundaries between meaningful segments in a satellite image, for instance, between buildings, roads or forested areas. The research results used in this patent were outlined by Burnaev and Bokhovkin in a conference paper, published in proceedings of International Symposium on Neural Networks (ISNN), also available as an arXiv preprint.

Such segmented maps have various applications in environmental monitoring, urban planning, forestry, agriculture, and other geospatial analysis. “For example, we can count cars in a parking lot in order to predict traffic, or measure the area of agricultural fields. The recognition of boundaries of objects is very important if we want to perform efficient change detection, for example, to detect an increase in the size of a landfill, or assess damage due to catastrophic events like hurricanes,” Burnaev said.

The team plans to extend their approach to handle multi-class semantic segmentation tasks – currently each object class needs a separate neural network.

Contact information:
Skoltech Communications
+7 (495) 280 14 81