“Applying AI to a new type of data isn't a huge step once you have a grasp on how AI systems work,” Stephanie says. Her friend had introduced her to Zhi Huang, a marine environmental modeller from Geoscience Australia, who was looking to fill gaps in satellite images.
When satellites take snapshots of the ocean, there are always at least some areas covered by clouds, and the current artificial intelligence-based models have not been accurate in terms of calculating what is beneath these clouds.
The research was a team effort, Stephanie explains. “Zhi Huang from Geoscience Australia, who invited me to help out, first explained the actual geoscience behind it to me. Taking that further and applying it with AI was reasonably easy,” Stephanie says.
“Zhi was also the one who collected the data from the satellite and did some initial pre-processing on it as well before he passed the data on to me. I was responsible for the development of the AI and a lot of the processing of the data.”
JCU’s Bronson Philippa, Senior Lecturer for Electronic Systems and IoT Engineering, was also on the team. “Bronson helped me with brainstorming ideas, including a novel time-based penalty that our final AI model used to help it fill gaps more accurately,” she says.
A new way of looking at sea surface data
Stephanie, Zhi and Bronson weren’t the first to conduct this kind of research, but the models used previously needed a lot of computational power and weren’t very accurate. “We were trying to obtain a lower error rate in terms of temperatures and structure of the currents,” Stephanie says.
The team experimented with several AI models for almost six months, and there were many disappointments. “Some conventional machine learning approaches, such as convolutional neural networks for image-based data, weren’t successful.
"These models are good at finding features within images, but we had no luck with those, as they weren’t accurate,” Stephanie says. “We also tried out more advanced models that are usually used for reconstruction tasks. They didn’t produce the results we were looking for either.”