The goal of this competition is to identify, segment, and classify deforestation drivers from satellite images. For the competition, the following four classes of deforestation are as follows; grassland/shrubland, logging, mining, and plantation.

The dataset consists of satellite images from the Sentinel-2 dataset, where each image contains 12 bands.

Link to competition: here

Our report can be found here.

Made by: Simon Vedaa, Khalil Ibrahim, and Safiya Mahamood

Make sure uv is installed.

Run: uv sync, to install all dependencies.

To train a model, run the following command:

python train.py --epochs 3 --config unet

Available configs:

• unet
• deeplab
• segformer
• transunet
• vit_seg

Model checkpoints are stored in ./data/training_result/config_name by default.

Both transunet and vit_seg configs can be run with additional parameters to freeze the encoders during training.

python train --epochs 3 --config transunet --frozen_start --f_epochs 1

Use frozen_start to freeze the encoder, and f_epochs determines for how many epochs to freeze.

To run model selection on validation data and create a submission:

python eval.py

The submission is stored in data/submission.json

• Models.py: Pytorch Lightning module containing training, validation, and testing loop/logic.
• train.py: Script for training models.
• eval.py: Script for model selection and creating competition submission.
• utils.py: Contains various utility functions.
• ensemble.py: Module for creating ensemble models.
• datasets.py: Module for defining dataset and Datamodule used in Models.py
• configs.py: Contains configurations for the models.
• generate_masks.py: Script for generating segmentation masks/labels from competition data.
• visualization.ipynb: Notebook for creating visualization of predictions.
• graphs.ipynb: Notebook for plotting training and validation metrics.
• report/: Contains the final report and all files and scripts to generate it.
• trans_unet/: Contains all code and documentation for our implementation of TransUNet.
• vision_transformer/: Contains all code and documentation for our implementation of Vision Transformer.
• segformer/: Documentation of our usage of Segformer.

• We used motokimuras baseline as a starting point for our solution: here