Southern Ocean sea ice has undergone rapid, unprecedented change over the past nearly two decades.
After a long-term small but steady increase in sea-ice extent, Antarctic winter extent ramped up, peaking above 20 Mkm2 in 2014 before dropping to summer (2016/17) and then winter lows (from 2023). The latter is consistent with an abrupt system shift. This behavior was not projected by state-ofthe- art numerical models, highlighting urgent gaps in our understanding of the processes that control Southern Ocean sea ice variability and change.
A key barrier to our understanding is the sparse and uneven distribution of observations in this remote and challenging environment. In situ measurements provide essential detail on ice, snow, and atmosphere-ocean-ice interactions, but their spatial and temporal coverage is inherently limited. Remote sensing observations offer broad spatial coverage but are constrained by resolution limits, complex and heterogeneous sea-ice surface conditions, limited sensitivity to subsurface processes, uncertainties in derived products, and a lack of in situ observation for validation. As the demand for reliable predictions grows, there is a pressing need for an observing framework that makes the most efficient use of complementary measurement systems.
The InSync campaign offers a unique opportunity to address this challenge by coordinating a three-year program of continuous observations, integrating both in situ and satellite measurements. To fully capitalize on this effort, the community would benefit from tools that quantitatively evaluate where new measurements will have the greatest impact. Historically, this type of assessment has relied on numerical modeling, or case-specific expert judgement. Statistical and model-based frameworks have also been used to identify optimal instrument placement (Ponsoni et al., 2020). Additionally, recent advances in AI have shown promise in determining potential paths-of-travel for saildrones that would provide improved skill in AI-based global reconstructions of pCO2 (Heimdal et al., 2024).
The OAISIS working group will build on these approaches to develop an open-source, AI-enabled framework for evaluating and comparing alternative sea ice observing strategies. Rather than focusing on traditional Observing System Simulation Experiment (OSSE)-style experiments, our approach emphasizes AI models trained on climate model output and satellite-derived data to identify sampling configurations that:
This effort will be closely coordinated with InSync’s observation-based cross-cutting groups (satellite, land-based, airborne, ship-based, instruments-based) to ensure that the practical constraints and opportunities associated with each observing platform are incorporated into the system design. By uniting AI and field expertise, OAISIS aims to deliver actionable, community-informed guidance that strengthens the Southern Ocean observing network and enhances our ability to understand and predict sea ice in a rapidly changing environment
The overarching scientific objective is to highlight critical locations to acquire observations to (i) improve the prediction skill of Antarctic SIE and (ii) enable full spatial reconstructions of sea ice thickness from sparse measurements. At future stages of the project, estimates of snow depth could complement the sea ice thickness reconstructions.
Specific: Develop an open-source, AI-enabled framework that evaluates and optimizes sea ice observing strategies using machine learning prediction and reconstruction models.
Measurable: Quantify information gain and uncertainty reduction associated with specific observing configurations using metrics such as forecast error reduction, reconstruction uncertainty, and sensitivity diagnostics, evaluated both prior to the InSync observation campaign (the basis for the design of our recommendations) and as new in situ measurements become available.
Achievable: Apply the framework to a range of observing platforms (e.g. satellite, ship-based, autonomous) to assess the relative impact of spatial distribution, temporal frequency, and (where feasible) adaptive sampling potential.
Relevant: Sea ice thickness and extent are central to understanding the freshwater cycle and sea ice mass balance in the Southern Ocean.
Time-bound: Deliver an initial set of AI-informed observing priorities (i.e., recommended sampling locations) and evaluation metrics by mid-year 2027, timed to inform the InSync observational phase.
The overarching coordination objective is to facilitate interaction between AI-based observing-system design and observation-focused InSync working groups and to integrate AI-driven insights into observational planning, while sharing our framework, methods, and expertise to strengthen the community’s capacity for sea ice observing system design. The working group will also serve as a forum for shaping and aligning proposal development to national and international funding calls, ensuring that our observing-system design activities are scientifically coherent and well integrated across institutions.
Specific: Establish collaboration channels (e.g. shared github repositories, focused technical exchanges) with major observation-focused teams to co-develop use cases, share methodological advances, and tailor the AI-based recommendations to their operational needs.
Measurable: Collect metrics of engagement including tracking use of github repository and attendance at training workshops. Ensure a diversity of representation across the InSync research community, including ECRs.
Achievable: Leverage existing project networks and InSync group structures, including a small observational “interface group” that connects OAISIS with observation-focused InSync WGs. Provide user-friendly documentation and support to ensure partners can adopt the AI framework without deep expertise in machine learning.
Relevant: Close coordination ensures that the AI-driven optimal sampling strategies are directly informed by real-world observation constraints and support the broader community’s data needs.
Time-bound: Formalize collaborations by Q2 2026; host AI training workshops and knowledge-exchange activities during 2026-2027
OAISIS will develop two complementary AI-based reference systems, designed explicitly as controlled testbeds to evaluate the impact of alternative observing strategies for Antarctic sea ice:
1. An Antarctic sea-ice extent (SIE) prediction system, designed to produce predictions of Antarctic SIE and to quantify how additional observations could improve forecast still. This model will build on existing SIE prediction approaches (i.e., with neural networks) and will be structured to quantify the marginal value of additional observations for forecast skill (e.g. through transfer-learning, perturbation experiments, explainable AI (XAI) diagnostics (Hoffman et al., 2025)). The primary outputs will include SIE forecasts, forecast uncertainty, and diagnostics identifying which variables, regions, and timescales most strongly influence predictive skill – thereby providing guidance on where additional observations would most effectively inform SIE prediction. This prediction system will feed established coordinated forecast intercomparison exercises such as SIPN South
(https://fmassonn.github.io/sipn-south.github.io/; Massonnet et al., 2023)
2. An AI-based reconstruction system for sea-ice thickness (SIT), designed to produce global reconstructions of Antarctic SIT from sparse and heterogeneous observations. Using a climatemodel / OSSE-like framework, this system will enable controlled experiments in which candidate observing configurations (e.g., spatial location, sampling frequency) are systematically introduced or withheld. The resulting changes in SIT reconstruction skill and uncertainty will be used to assess where additional observations would most improve SIT reconstructions. This system will leverage existing ML reconstruction approaches developed for sea-ice concentration and related geophysical fields (Heimdal et al., 2024). The primary outputs will include full spatial SIT reconstructions with associated uncertainty estimates, along with diagnostics identifying which variables, regions, and timescales most strongly influence reconstruction skill – thereby informing observing priorties.
While OAISIS targets SIE prediction & SIT reconstruction directly, the framework will be designed to infer which additional sea ice, oceanic, and atmospheric variables are most influential for achieving skill in prediction and reconstruction. The methodological approach builds on established numerical, statistical, and machine-learning-based observation-system design concepts in sea ice environments (Massonnet, 2019), with emphasis on pre-campaign observing-system evaluation.
Products: Our main contribution will be an open-source AI framework for evaluating Southern Ocean sea ice observing strategies. Along with this, we will develop machine learning models that (i) forecast SIE while incorporating recommended sampling strategies (using transfer learning) and (ii) make global reconstructions of SIT.
Datasets: Pre-Obs: Demonstration datasets illustrating global reconstructions of SIT from sparse observations, with uncertainty estimates. Post-Obs (conditional on data availability): New data product(s) of multi-scale sea-ice thickness distributions at sub-seasonal resolution.
Publications: Concept and methods paper/s describing the AI framework/s; follow-up publication/s demonstrating improvements in sea ice prediction and SIT reconstruction skill with collected observations.
Training Workshops: Host multi-site and online workshops for targeted training in AI methods for polar science, as well as training in the use and interpretation of the OAISIS framework so that it can be applied effectively.
Capacity Building Activities: We plan to hold regular (approximately quarterly) working group meetings online and further focused interactions with the different InSync observation-based groups (land-based, flight-based, EO, ship-based).
| Timeframe | Task |
|---|---|
| Q4 2025 | Communicate to potential collaborators |
| Q1 2026 | Submit working group application (early January) |
| Q2 2026 | Pilot OAISIS framework and diagnostics |
| Q3-Q4 2026 | Fine-tune OAISIS models and evaluation metrics |
| 2027-2030 | Implementation of OAISIS with observations |
| 2030 onward | Analysis of data and validation of methods |
The OAISIS group will work in collaboration with other Antarctica InSync working groups and thematic groups. We aim to start in collaboration with the sea ice thematic group (theme 2), but recognize that once our methods are developed they could be applicable to other thematic groups such as the ocean group (theme 1). We plan close collaboration with the InSync observation-based cross-cutting groups (land-based, flight-based, EO, ship-based) in order to coordinate our recommendations for observing system design with platform-specific constraints and InSync’s broader observational strategy.
