| UN SDGs | 13 Climate Action |
|---|---|
| Impact | United Kingdom |
| Category | Artificial Intelligence |
| Tags | #AI #icebergs #Satellites |
AI has been trained to measure changes in icebergs 10,000 times faster than a human could do it.
This will help scientists understand how much meltwater icebergs release into the ocean – a process accelerating as climate change warms the atmosphere.
Scientists at the University of Leeds in the United Kingdom say their AI can map large Antarctic icebergs in satellite images in just one-hundredth of a second, reports the European Space Agency.
For humans, this task is lengthy and time-consuming, and it’s hard to identify icebergs amid the white of clouds and sea ice.
In a major breakthrough, researchers from the University of Leeds have developed a neural network that can quickly and precisely map the extent of large Antarctic icebergs in satellite images. This new artificial intelligence approach can accomplish the task in just 0.01 seconds, vastly faster than the tedious manual analysis previously required.
Lead author Anne Braakmann-Folgmann, who conducted the research during her PhD studies at Leeds, explained that accurately tracking icebergs is crucial since they impact ocean physics, chemistry, and biology. “It is vital to locate icebergs and monitor their extent to quantify their meltwater contribution to the oceans,” she said.
The Copernicus Sentinel-1 radar satellite plays a key role by providing images regardless of cloud cover and darkness. While icebergs, sea ice, and clouds appear indistinguishable in camera images, in radar images icebergs stand out brightly against the darker background. However, even in radar images, differentiation can be challenging when backgrounds are complex.
The neural network’s strength lies in its ability to grasp subtle relationships and contextual information across entire images. It reliably identifies even smaller iceberg fragments and excels at pinpointing the largest iceberg, which is essential for continuous monitoring.
The system was trained on manually-outlined Sentinel-1 images showing icebergs in diverse conditions. By continually adjusting its parameters during training, it learned to replicate the human-derived outlines with over 99% accuracy. Testing on icebergs spanning 54 to 1052 square kilometers demonstrated impressive real-world performance.
According to Dr. Braakmann-Folgmann, this automation enables easier observation of iceberg area changes. ESA’s Mark Drinkwater added that the machine learning approach provides an accurate and robust method to monitor this vulnerable region.