Satellite radar images of sea ice are so full of visual noise that even artificial intelligence struggles to tell ice from open water. The problem is that existing AI training datasets rely on people manually labelling radar images, which is slow and expensive. As a result, those datasets cover only small parts of the Arctic during specific seasons, leaving the Antarctic and winter months largely unlabelled. This project will automatically generate a far larger, more diverse dataset by pairing radar images from the Sentinel-1 satellite with visible-light images from the MODIS satellite taken within one hour of each other. The visible images, which clearly show ice in cloud-free daylight, will be used to create binary ice-water labels that can train AI models. If successful, the open-source dataset and codebase could improve AI-driven climate monitoring, ship navigation safety in polar waters, and habitat tracking for Arctic and Antarctic wildlife. The initial analysis suggests that for 2020 alone, over 10,000 high-quality image pairs could be identified for the Antarctic, far exceeding the scale of manually annotated datasets.
View original technical description
Objective: Generate an AI-ready dataset of sea ice from satellite imagery. Rationale: Understanding where sea ice is located and how it is changing over time is crucial for monitoring the effects of climate change in the polar regions, identifying suitable habitat locations for important Arctic and Antarctic wildlife, and ensuring the safe navigation of ships in areas of the ocean covered by sea ice. One powerful dataset for tracking sea ice is radar satellite imagery, captured from the Sentinel-1 satellite, which is able to ‘see’ through cloud and darkness to detect sea ice all year round and in all weather conditions. However, interpreting radar imagery is not straightforward. Unlike visible images, radar images contain a lot of noise or speckle, sometimes making it difficult to automate systems to identify where sea ice is located. To train Artificial Intelligence (AI) systems to detect sea ice, we require large, high-quality datasets where radar imagery is paired with labels that can inform the system of where sea ice is present. Currently, most labelled data is generated from manually annotating the image. This approach is time-consuming, meaning most of the labelled data is limited to very restricted parts of the Arctic and is only produced for radar images captured during particular times of the year. This limits the usefulness of this labelled data for training AI models to identify sea ice all year round and in both hemispheres. This project will develop a large, publicly available, AI-ready dataset of sea ice coverage, automatically derived from satellite images. We will identify pairs of overlapping radar and visible satellite images, captured from the MODIS satellite, that have been acquired in the same place at the same time. The visible images will be used to distinguish between ice and water in cloud-free, daylight conditions. This will produce a triplet dataset contained labelled radar and visible satellite imagery that can be used for training AI models and for other environmental and operational applications. Our project contains three main steps: · Match radar and optical satellite images: We will identify radar images that have been captured within one hour of a visible image from the same location. This short time window is essential due to the very dynamic nature of sea ice that can drift tens of kilometres each day. · Generate sea ice labels: Automated methods will be applied to the visible images to produce binary masks that clearly distinguish between ice and water in each image. These masks will be paired with the original radar and visible satellite images. · Build a reusable codebase for the future: We will publish an open-source codebase alongside the dataset, enabling other users to apply the same methods and generate additional data as new satellite imagery is made available. Our initial analysis shows that in just one year (2020), over 10,000 high-quality patches of radar-optical image pairs could be identified for the Antarctic alone. Scaling this up to include data from multiple years and both polar regions could result in a dataset that is far larger and more diverse than pre-existing datasets produced through manual annotation. This will increase the range of downstream applications for this dataset from supporting AI-driven climate research, to increasing navigation safety, and improving environmental monitoring.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know