Kelp Wanted: Segmenting Kelp Forests

General

• Competition Purpose: Help researchers estimate the extent of Giant Kelp Forests by segmenting Landsat imagery.
• Type: Semantic Segmentation
• Host: Mathworks
• Platform: Drivendata
• Competition link: https://www.drivendata.org/competitions/255/kelp-forest-segmentation/
• Placement: 3rd (3/671)
• User Name: Ouranos
• System used: GPU Server mainly and Google Colab Pro (for 2 models of high image size)
• Solution: write-up
• Train duration: About 3 days locally and 1 day on Google Colab pro
• Inference duration: less than 1 hour depending on hardware

High Level solution description. The solution is based on the ensemble of 9 segmentation models predictions. Each one of these predictions is a simple average of all possible flips (TTA). At post processing, a sea-land mask derived from the last image channel (altitude) was applied. Finally, probabilities turned to mask using a 0.43 threshold.

System

Characteristic	Main Server	Google Colab pro
Operating System	Ubuntu 18.04	Ubuntu 22.04
Cuda	11.2	12.2
Python	3.8	3.10
Pytorch	1.8	2.1
GPU	GeForce GTX 1080	Tesla V100-SXM2-16GB
RAM	70GB	50GB

Create virtual Environment

virtualenv -p python3.8 kelpvenv

Activate Environment

source kelpvenv/bin/activate

Install the appropriate pytorch version for the installed cuda version

install the rest libraries with pip or pip3

pip install -r requirement.txt

Configuration

Modify the code/config.py to specify the paths. Trained model weights should be placed in trained_weights directory.
Set RETRAIN to True to retrain all models.

Trained models

Extract trained models weights from zip file and set trained_weights directory path to config.py file.

Generate submission

First run all jupyter notebooks starting with 'segm-'.
This will generate equal number of probability prediction files in test_preds directory.
Then run ensemble-postprocess-submission.ipynb to combine all predictions and make the final predictions masks and submission file.

Re-train models

Set RETRAIN to True and run all notebooks. This will save new models weights in weights directory and will overwrite existing test predictions in test_preds directory unless a new test_preds directory is provided in the config file.

Regenerate submission

Edit the code/config.py file to change weights path (use new weights). Then rerun all jupyter notebooks as before.

Extra notebook

There is an extra notebook provided (kelp-uppernet-512-full.ipynb). This can be used to replicate an alternative solution, scoring as described/proposed solution. To generate this solution after running this notebook too, uncomment and run the last cell in ensemble-postprocess-submission.ipynb.

Citing This Work

This code has been used in the research paper "Enhancing kelp forest detection in remote sensing images using crowdsourced labels with Mixed Vision Transformers and ConvNeXt segmentation models". If you find this code useful, please consider citing it.
This reasearch was published by International Journal of Remote Sensing and It's preprint can be found on arxiv.

If you use this repository in your research, please cite the following paper:

BibTeX:

@article{nasios2025enhancing,
  author = {Ioannis Nasios},
  title = {Enhancing kelp forest detection in remote sensing images using crowdsourced labels with Mixed Vision Transformers and ConvNeXt segmentation models},
  journal = {International Journal of Remote Sensing},
  volume = {46},
  number = {5},
  pages = {2159--2177},
  year = {2025},
  publisher = {Taylor \& Francis},
  doi = {10.1080/01431161.2024.2448307},
}