Datasets
Standard Dataset
pan-Arctic region 60-day forecast data and remote sensing ground truth
- Citation Author(s):
- Submitted by:
- Jianxin He
- Last updated:
- Mon, 12/30/2024 - 10:17
- DOI:
- 10.21227/vnp3-6396
- Data Format:
- License:
5 Views- Categories:
- Keywords:
Abstract
output mat: 60-day SR-SICFormer forecasting data
label mat: 60-day remote sensing ground truth
nan_mat: land-sea mask
maks: lat-lon mask
This study proposes a data-driven SR-SICFormer model combined with a ConIce loss function, aiming to address these issues mentioned above. SR-SICFormer is a novel Transformer-based model that combines image super-resolution reconstruction with spatiotemporal feature learning, enabling high-resolution daily SIC predictions for the next 15 days to 2 months. The ConIce loss function incorporates SIC conservation, correcting SIE and internal concentration predictions by integrating physical constraints from thermodynamic and dynamic sea ice processes. Compared to existing models, this approach performs excellently in fine-scale, long-term, extended-range daily SIC prediction. Furthermore, for the 60-day forecast task during the melting season, SR-SICFormer continues to perform exceptionally well.
This study proposes a data-driven SR-SICFormer model combined with a ConIce loss function, aiming to address these issues mentioned above. SR-SICFormer is a novel Transformer-based model that combines image super-resolution reconstruction with spatiotemporal feature learning, enabling high-resolution daily SIC predictions for the next 15 days to 2 months. The ConIce loss function incorporates SIC conservation, correcting SIE and internal concentration predictions by integrating physical constraints from thermodynamic and dynamic sea ice processes. Compared to existing models, this approach performs excellently in fine-scale, long-term, extended-range daily SIC prediction. Furthermore, for the 60-day forecast task during the melting season, SR-SICFormer continues to perform exceptionally well.
