🚀 Frequency-Enhanced Hilbert Scanning Mamba for Short-Term Arctic Sea Ice Concentration Prediction, IEEE TGRS 2026

Model

FH-Mamba for Predicting Sea Ice Concentration (SIC) in the Arctic.

This study utilizes Frequency-enhanced Hilbert scanning Mamba to simulate the spatial-temporal evolution of sea ice concentration. The model incorporates Discrete Wavelet Transform (DWT) for frequency information extraction and Hilbert Scanning 3D Mamba for capturing long-range spatial-temporal dependencies efficiently.

How to Use

Environment

Install the required dependencies (such as PyTorch, Mamba-ssm, etc.) to create the python environment.

Dataset

We recommend downloading and organizing the dataset in the following order to avoid any issues.

1. Download and Reorganize Data

The monthly SIC data used in this study can be downloaded from OSI SAF OSI-450-a1 (Global Sea Ice Concentration Climate Data Record v3.0), available at https://osi-saf.eumetsat.int/products/osi-450-a1, which also contains a detailed description of the dataset and user guide.

Run the download.py file in the data directory to download the data.

Run organize.py in the data directory to reorganize the data.

The reorganized data structure:

data
├── 1991
├── 1992
├── 1993
......
├── 2019
│   ├── 01
│   │   ├── ice_conc_nh_ease2-250_cdr-v3p1_201901011200.nc
│   │   ├── ice_conc_nh_ease2-250_cdr-v3p1_201901021200.nc
│   │   ├── ice_conc_nh_ease2-250_cdr-v3p1_201901031200.nc
│   │   ├── ice_conc_nh_ease2-250_cdr-v3p1_201901041200.nc
......
├── 2020

2. Generate file containing the path to all the data files

bash gen_data_text.sh

The generated data.txt file is located in the data directory.

The content of data.txt should be as follows:

./1991/01/ice_conc_nh_ease2-250_cdr-v3p1_199101011200.nc
./1991/01/ice_conc_nh_ease2-250_cdr-v3p1_199101021200.nc
./1991/01/ice_conc_nh_ease2-250_cdr-v3p1_199101031200.nc
./1991/01/ice_conc_nh_ease2-250_cdr-v3p1_199101041200.nc
......

3. Read and store the processed data to facilitate future data loading

Run full_sic.py. This will call the write_netcdf function, passing the filename from the previous step as an argument, to read and store the processed data. A full_sic.nc file will be generated in the data directory to facilitate future data reading.

# full_sic.py
# Read and store the processed data for easier loading next time by calling the write_netcdf function 
# and passing the output filename from the previous step as an argument.
from utils import write_netcdf
start_time = 19910101
end_time = 20201231
write_netcdf("data.txt", start_time, end_time, "full_sic.nc")

Train

Change relevant parameters and the .nc file path in the config.py file, put all the scripts under the same folder and run:

python train.py

The training progress will be printed to the console. You can also choose to redirect this information to a log file.

Test

Specify the testing period and output directory, then run:

python test.py -ts 20160101 -te 20161231
python test.py -ts 20170101 -te 20171231
python test.py -ts 20180101 -te 20181231
python test.py -ts 20190101 -te 20191231
python test.py -ts 20200101 -te 20201231

Args:

parser.add_argument('-ts', '--start_time', type=int,
                        required=True, help="Starting time (eight digits, YYYYMMDD)")
parser.add_argument('-te', '--end_time', type=int,
                        required=True, help="Ending time (eight digits, YYYYMMDD)")