AI4ER Guided Team Challenge 2023:
Sea Ice Classification

This repository contains code for the Sea Ice Classification Challenge from the 2022-23 cohort of AI4ER.

AI4ER is the the UKRI Centre for Doctoral Training (CDT) in the Application of Artificial Intelligence to the study of Environmental Risks at the University of Cambridge.

Project Description

This goal of this 3-month project was to automatically classify sea ice concentration in the East Weddell Sea, Antarctica. The Weddell Sea is an active area of iceberg calving [1] and a critical shipping route for access to the Halley Research Station, a research facility operated by the British Antarctic Survey. Accurate assessments of sea ice concenration in the East Weddell Sea are hence of great importance to ensure the safety and success of future expeditions.

Data

This project uses two publicly available datasets:

• Labeled sea ice charts jointly developed by the Arctic and Antarctic Research Institute, USA National/Naval Ice Center, and Norwegian Meteorological Institute.
• Sentinel 1A and 1B Synthetic Aperture Radar (SAR) satellite imagery provided by the Copernicus Open Access Hub, operated by the European Space Agency. The Copernicus Open Data Policy enables free, open access to Sentinel products [2]. Sentinel Terms and Conditions can be found at the following link.

Region of Interest

Relative locations and areal extents of (b) training, (c) validating, and (d) testing subsets.

Temporal and geographical coverage of the dataset.

Pair Number	Subset	Date of Sea Ice Concentration Chart	Observation Window of SAR signals	Geographical Coverage
1	training	Nov 6, 2017	10:25:50 PM - 10:26:55 PM	Weddell Sea
2	training	Dec 23, 2017	10:42:00 PM - 10:43:04 PM	Weddell Sea
3	training	Jan 4, 2018	10:41:59 PM - 10:43:03 PM	Weddell Sea
4	training	Feb 22, 2018	11:22:34 PM - 11:23:38 PM	Weddell Sea
5	training	Feb 23, 2018	12:59:54 AM - 1:00:58 AM	Weddell Sea
6	training	Feb 26, 2018	1:24:13 AM - 1:25:17 AM	Weddell Sea
7	training	Dec 3, 2018	8:18:15 AM - 8:19:14 AM	Antarctica Peninsula
8	training	Dec 9, 2018	8:51:00 AM - 8:52:08 AM	Antarctica Peninsula
9	training	Dec 10, 2018	11:50:06 PM - 11:51:05 PM	Antarctica Peninsula
10	training	Dec 20, 2018	8:19:15 AM - 8:20:19 AM	Antarctica Peninsula
11	training	Mar 5, 2020	12:35:47 AM - 12:36:52 AM	Weddell Sea
12	training	Mar 13, 2020	1:08:15 AM - 1:09:20 AM	Weddell Sea
13	training	Dec 23, 2021	11:14:52 PM - 11:15:43 PM	Weddell Sea
14	validating	Mar 13, 2019	11:22:41 PM - 11:23:45 PM	Weddell Sea
15	validating	Jan 17, 2020	10:01:39 PM - 10:02:43 PM	Weddell Sea
16	testing	Dec 16, 2022	11:23:48 PM - 11:24:53 PM	Weddell Sea
17	testing	Dec 22, 2022	1:09:23 AM - 1:10:27 AM	Weddell Sea
18	testing	Jan 12, 2023	8:32:41 PM - 8:33:45 PM	Weddell Sea

Sample visualizations of (a) water and ice, (b) HH band, (c) HV band, (d) angle, and (e) HH/HV.

Categories of sea ice concentrations.

Code	Description	Code	Description	Code	Description
0	Ice Free	35	3/10 to 5/10 ice	78	7/10 to 8/10 ice
1	Open Water ( extless 1/10 ice)	40	4/10 ice	79	7/10 to 9/10 ice
2	Bergy Water	45	4/10 to 5/10 ice	80	8/10 ice
10	1/10 ice	46	4/10 to 6/10 ice	81	8/10 to 10/10
12	1/10 to 2/10 ice	47	4/10 to 7/10 ice	89	8/10 to 9/10 ice
13	1/10 to 3/10 ice	50	5/10 ice	90	9/10 ice
14	1/10 to 4/10 ice	56	5/10 to 6/10 ice	91	9/10 to 10/10 ice, 9+/10 ice
20	2/10 ice	57	5/10 to 7/10 ice	92	10/10 ice
23	2/10 to 3/10 ice	60	6/10 ice	99	Unknown/Undetermined
24	2/10 to 4/10 ice	67	6/10 to 7/10 ice	-9	Null Value
30	3/10 ice	68	6/10 to 8/10 ice	100	Full Ice
34	3/10 to 4/10 ice	70	7/10 ice

Schematic illustration of the overlapping tiling technique.

Models

This project uses three models:

1. A baseline Decision Tree (DT) model
2. A basic U-Net
3. A pretrained resnet34 from the segmentation_models_pytorch Python library, which is distributed under the MIT license.

Code Structure

|──────tutorial                     <-- Folder containing tutorial CNN notebook
|──────info.md                      <-- Sea ice concentration values
|──────JASMIN.md                    <-- Step-by-step guide
|──────constants.py                 <-- SAR/ice chart pairs, binary/ternary classes
|──────environment.yml              <-- List of Python modules
|──────interesting_images.csv       <-- List of tiles containing all three categories for ternary classification
|──────interesting_images.py        <-- Generate interesting_images.csv
|──────intro_deep_learning.ipynb    <-- Deep learning notebook
|──────metrics.csv                  <-- Lists mean and std dev of all SAR images
|──────metrics.py                   <-- Calculate metrics.csv
|──────metrics_per_pair.csv         <-- Lists mean and std dev for individual SAR image
│──────model.py                     <-- Unet model and evaluation metrics
│──────split.py                     <-- Construct training & validation datasets
│──────test.py                      <-- Test CNN and save output to WANDB
|──────test_scikit.py               <-- Test DT and save output to WANDB
|──────test_slurm_script.py         <-- Test the model on JASMIN
|──────test_slurm_script_scikit.py  <-- Test the DT on JASMIN
|──────tiling.py                    <-- Generate tiles from SAR/ice chart paris
│──────train.py                     <-- Train CNN and save output to WANDB
|──────train_scikit.py              <-- Train DT and save output to WANDB
|──────train_slurm_script.sh        <-- Train the model on JASMIN
|──────train_slurm_script_scikit.sh <-- Train the DT on JASMIN
│──────util.py                      <-- Load data into the model and normalise
|──────util_scikit.py               <-- Load data into the model, normalise and create the training dataset

Workflow

CNN Workflow (U-Net and ResNet34)

Decision Tree Workflow

Usage

To test the code and recreate the results of this project, follow the steps below:

1. Clone this repository (for the latest version) or retrieve the archived copy from Zenodo

2. Create and activate the conda environment using conda activate environment.yml, which contains all required Python modules and versions.

3. To generate ice chart and SAR tile pairs of 256x256 dimensions run: python tiling.py. Tile pairs containing NaN values will be discarded.

4. Follow the steps in JASMIN.md to train and test the CNN or DT model. Input arguments that were modified for this project include:

   Argument	Options	Default
   --model	unet, resnet34	unet
   --classification_type	binary, ternary	binary
   --sar_band3	angle, ratio	angle

Contributors

Project core members contributed equally to this work:

• Joshua Dimasaka, AI4ER MRes Student (2022 Cohort), University of Cambridge
• Andrew McDonald, AI4ER MRes Student (2022 Cohort), Univeristy of Cambridge
• Meghan Plumridge, AI4ER MRes Student (2022 Cohort), University of Cambridge
• Jay Torry, AI4ER MRes Student (2022 Cohort), University of Cambridge
• Andrés Camilo Zúñiga González, AI4ER MRes Student (2022 Cohort), University of Cambridge

With special thanks to our advisors for their project guidance and technical support:

• Madeline Lisaius, AI4ER PhD Student (2021 Cohort), University of Cambridge
• Jonathan Roberts, AI4ER PhD Student (2021 Cohort), University of Cambridge
• Martin Rogers, AI-lab, British Antarctic Survey

References

[1] Brunt ice shelf in Antarctica calves giant iceberg (2023) British Antarctic Survey. Available at: https://www.bas.ac.uk/media-post/brunt-ice-shelf-in-antarctica-calves-giant-iceberg/ (Accessed: March 9, 2023).

[2] The European Space Agency (n.d.) FAQ content, FAQ - Sentinel Online - Sentinel Online. Available at: https://sentinel.esa.int/web/sentinel/faq (Accessed: March 17, 2023).

Significant parts of our deep learning code are inspired by / adapted from the following:

https://github.com/ampersandmcd/DeepExtremeMixtureModel

https://github.com/ampersandmcd/COMETFlows

https://github.com/ampersandmcd/CSE803FinalProject