Sea Ice Extent Forecasting using Convolutional LSTMs

Abstract

The Arctic sea ice cover is a crucial component of the Earth’s climate and ecosystem. Climate change has an especially pronounced effect on the Arctic, making it difficult to predict subseasonal-to-seasonal Arctic sea ice extent. In this paper, a data-driven approach is taken to predict monthly Arctic sea ice extent using deep learning. We first assess the work of (Ali et al.) in their paper "Sea Ice Forecasting using Attention-based Ensemble LSTM" and find a lack of reproducibility in their methodology. Moreover, we show that naive statistical baselines outperform their proposed model. We then propose an alternative model architecture which accepts both spatial and temporal inputs, and uses several new climatic input variables. We test a variety of hyperparameters with our best model achieving a %RMSE of 4.88% and MAPE of 3.29%, outperforming a climatology with linear trend baseline by 56% in %RMSE.

Project Navigation

This project consists of 1) replicating Ali et al.'s original results, 2) enhancing their architecture using their data, 3) evaluate traditional ML methods on the original dataset, and 4) proposing and evaluating a novel model architecture using new data. Work done to fulfill (1) and (2) can be found in folders named 'Replication', (3) can be found in 'ML_Models', and (4) can be found in 'Improvements'.

The final scores of our proposed models are in Results/Improvements/all_results.csv, and visualizations can be found under Models/Improvements/evaluation.ipynb. Note that the 'Improvements' notebooks cannot be run without the original spatial data, which is too large to upload to Github.

Folder Structure

.
├── Data
│   ├── Arctic_domain_mean_1979_2018.csv             ┐
│   ├── Arctic_domain_mean_monthly_1979_2018.csv     │
│   ├── dailyt30_features.npy                        │- Original data from Ali et al.  
│   ├── dailyt30_target.npy                          │ 
│   ├── monthly_features.npy                         │
│   ├── monthly_target.npy                           ┘ 
│   └── extents.nc  <-- SIE calculated from ERA5 SIC estimates
├── Models
│   ├── Improvements
│   │   ├── model.py  <-- Contains the main model class 
│   │   ├── pre-process.ipynb  <-- Pre-processing the spatial data & calculating SIE
│   │   ├── run_tests.ipynb  <-- Train models given a set of hyperparameters
│   │   ├── evaluation.ipynb  <-- Model evaluation and visualization
│   │   └── xgrid_utils.py  <-- Used to calcualte SIE
│   ├── ML_Models
│   │   └── ML_models.ipynb
│   └── Replication
│       ├── d-LSTM_Replication.ipynb  <-- Replicating the daily-to-monthly model results 
│       ├── m-LSTM_Replication.ipynb  <-- Replicating the monthly-to-monthly model results
│       ├── E-LSTM_Replication.ipynb  <-- Replicating the ensemble model results
│       ├── EA-LSTM_Replication.ipynb  <-- Replicating the attention-based ensemble model results
│       ├── EA-LSTM_Enhanced.ipynb  <-- Enhancing the original architecture 
│       └── EA-LSTM_Enhanced_Separated.ipynb  <-- Training each model separately 
└── Results
    ├── Improvements
    │   ├── all_results.csv  <-- Results of all experiments
    │   └── sensitivity_analysis.csv  <-- Results of the sensitivity analysis 
    ├── ML_Models  <-- CSV files containing scores for each ML model
    │   └── ...
    └── Replication  <-- Results of running the original models 20 times
        ├── d-lstm_scores.csv
        ├── ea_lstm_scores.csv
        ├── e_lstm_scores.csv
        └── m-lstm_scores.csv

General Model Architecture

Our proposed model architecture consists of stacked ConvLSTM, convolutional/pooling layers, and LSTM layers.

Best Model Architecture

References

Ali, S., Huang, Y., Huang, X., and Wang, J. Sea Ice Forecasting using Attention-based Ensemble LSTM. arXiv:2108.00853 [physics], February 2022. URL http://arxiv.org/abs/2108.00853. arXiv: 2108.00853.