A new artificial intelligence (AI) program is revolutionizing the way scientists analyze satellite and drone data. This program, known as METEOR, has the ability to train neural networks to characterize objects with just a handful of images.
Satellite and drone images provide valuable information about the Earth’s surface, including changes in animal populations, vegetation, ocean debris, and more. Neural networks can analyze these images at lightning speed and classify individual objects. However, existing AI programs struggle to switch from recognizing one type of object to another without extensive training on new data.
METEOR aims to change that. Developed by scientists from Ecole Polytechnique Federale de Lausanne, Wageningen University, MIT, Yale, and the Jülich Research Center, this chameleon-like application can train algorithms to recognize new objects after being shown only a few images. The findings of their study have been published in the journal Communications Earth & Environment.
Neural networks rely on annotated data to classify images accurately. The more data they are fed, the more precise their results become. However, in the field of environmental science, obtaining large datasets can be challenging. Specific phenomena or objects may be limited in number or dispersed, making it difficult to train AI programs effectively.
METEOR overcomes these challenges by utilizing adaptive algorithms and meta-learning. It can generalize results from previous deployments and apply them to new situations, significantly reducing the number of training images required. In fact, METEOR can deliver reliable results with just four or five high-quality images of an object.
To test the application, the developers modified a neural network trained to classify land occupation globally. They successfully used METEOR to carry out five recognition tasks: measuring vegetation coverage in Australia, identifying deforestation zones in Brazil, pinpointing changes in Beirut after the 2020 explosion, spotting ocean debris, and classifying urban areas into different land use categories.
The results showed that METEOR’s performance with a small dataset was comparable to AI programs trained for more extended periods with larger datasets. The researchers aim to further develop the application by training it on numerous tasks and integrating a user interface for human interaction.
METEOR’s ability to quickly characterize objects with minimal data has the potential to revolutionize environmental science and Earth observation. By harnessing the power of AI and satellite data, scientists can gain valuable insights into various phenomena, ensuring a more sustainable future for our planet.
