- Researchers from the University of California have built a machine learning model that uses satellite images of an area to create a real ground-level representation of that (including objects in the image)
- The model was trained on 16,000 pairs of images (ground level and aerial view); the technique used to create this model is called conditional generative adversarial network (cGAN)
- The final model produced accurate results 73 percent of the time, which is extremely impressive given the scope of the research
Remember our ‘Eye in the Sky’ article about using machine learning with drones to detect violence in a crowd of people? Well, that kind of technology has been in existence since decades, in the form of satellites. But only in the last few years has it been truly commercially used. It’s no coincidence that this has happened with the rise of machine learning in the industry.
Now, the latest study comes from a group of researchers at the University of California. They have trained a machine learning model that looks at satellite images of the Earth, and then creates a ground-level image. This is basically imagining what that particular area (in the image) actually looks like on Earth. This can potentially be used for mapping and creating images of the places which are out of our reach.
They have even published their research paper which you can read here. In it, the team has described the complete process of training the model and creating images. They used a machine learning system known as a conditional generative adversarial network (cGAN) which combines two neural networks: a ‘generator’ and a ‘discriminator’.
The cGAN was shown more than 16,000 pairs of images (comprising overhead images and ground-level images of the same location). It was trained to ‘visualise’ what objects looked like on the ground based on the photographs which showed an aerial shot of that area.
After training the system, the ‘generator’ neural net was fed with 4,000 new satellite images. Based on the training data, it was able to create fake ground-level images for these 4,000 pictures. Note that the ‘discriminator’ has access to the real ground-level views. By using the feedback from the discriminator, the generator gradually learns to produce more accurate images.
The fake images generated by the machine learning model looked similar to the original images but lacked depth in the form of more granular details. As you can see in the above image, the system is able to cover major details, like whether the image has road, land, or water, but some tiny details are missing from them.
Compared with the currently used human interpolation method (which is correct 65% of the times), this technique provided better results (correct 73% of the time). The researchers are determined to improve the system’s performance even more, and explore other machine-learning methods.
Our take on this
How cool is this? While most of the industry is looking at products and how to leverage machine learning to sell them, here comes a research totally isolated from that corporate environment. It’s such a refreshing change of pace and encourages data scientists to try their hand in these fields as well.
Go through the research paper and try to understand the way the model was designed. It will add a whole new perspective to way we usually go about building a model. Once done, check out the wonderful work Stanford is doing with satellite images and machine learning. It’s also open source and the code is available on their GitHub page so you don’t even need to search for ways to get started – it’s all here!
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!