Images taken by drones and satellites are invaluable tools for scientists, providing critical information about the Earth’s surface and its various changes. These images allow experts to train artificial intelligence (AI) programs to analyze and classify objects at an impressive speed. However, current AI programs have limitations when it comes to recognizing different types of objects in satellite images. Researchers from EPFL’s Environmental Computational Science and Earth Observation Laboratory, together with scientists from Wageningen University, MIT, Yale, and the Jülich Research Center, have set out to address this challenge.
Their solution is METEOR, a chameleon AI application that can train algorithms to identify new objects after being shown only a few images. Typically, training AI programs requires vast amounts of field data for each new object type. However, METEOR takes a different approach, using novel algorithms and methods that allow neural networks to generalize the results of previous deployments and apply that adaptation to new situations. This means that just four or five high-quality images are enough to retrain the system for a new task.
One of the difficulties in training neural networks on aerial and satellite images is the wide range of resolutions and spectral bands they encounter. METEOR overcomes this problem by being adaptable and capable of meta-learning. It learns from successful tasks it has previously solved and applies that knowledge to new situations. The researchers found that METEOR delivered reliable results for various recognition tasks using only a small dataset, comparable to AI programs trained for longer periods with more data. The program’s adaptability will be further perfected by training it on a multitude of tasks.
The team tested METEOR by modifying a neural network trained to classify different types of land occupation worldwide. They successfully used the adapted network to measure vegetation coverage in Australia, identify deforestation zones in Brazil’s tropical forest, assess changes in Beirut after the 2020 explosion, spot ocean debris, and classify urban areas based on land use types. The developers also plan to incorporate a user interface, allowing human users to click on high-quality images recommended by the neural network program.
Since the program will be shown only a few images, the relevance of those images is really important, says Marc Rußwurm, an assistant professor at Wageningen University and previously a postdoc at EPFL.
The groundbreaking potential of METEOR lies in its ability to provide accurate object recognition and classification from minimal data, which is particularly crucial in environmental science. It opens doors for studying region-specific phenomena and detecting statistically small but widely dispersed objects, such as ocean debris. This AI application has the power to revolutionize the field by facilitating faster and more efficient analysis of satellite images, ultimately enhancing our understanding of our changing planet.
In conclusion, METEOR’s chameleon-like abilities to recognize and classify objects in satellite images after being shown just a few examples offer promise for the field of environmental science. With the potential to revolutionize the analysis process and provide valuable insights into various phenomena, this innovative AI program marks a significant step forward in our ability to understand and monitor changes occurring on Earth’s surface.