This repository contains scripts that can be used to train a customized multi-task DeepLabV3 model to map sea ice parameters including Sea Ice Concentration (SIC), Stage of Development (SoD) and floe Size (FLOE) using the AutoICE challenge data.
Model architecture. A truncated ResNet-152 is used as backbone, followed by customized ASPP module with average pooling (instead of the default global pooling) to allow for training on sub-images and inference on full-scenes. One ASPP decoder module is used per each target of SIC, SoD, and Floe.
Please cite us!
The code in this repository was the main source for a sudy published as a peer-reviewed paper titled Enhancing and Interpreting Deep Learning for Sea Ice Charting using the AutoICE Benchmark. (https://doi.org/10.1016/j.rsase.2025.101538) The preprint is also available.
Clone this repository to your local machine using your tool of choice. Open the Anaconda Prompt (requires a working Anaconda installation):
Then, use the prompt to navigate to the location of the cloned repository. Install the environment using the command:
conda env create -f env_exported.yml
Follow the instructions to activate the new environment:
conda activate sea-ice-segment
We have two environment files:
- env_exported: the environment exported from Anaconda's history. This should be enough to replicate the results.
- env_full: the full environment installed. This includes more information and might be OS dependent. The experiments were executed using Windows 10 Pro for Workstations, mostly using version 21H2.
Use main.py to train the model according to configuration file. Evaluate the model using test.py that also requires configuration file