Context-Self Contrastive Pretraining for Crop Type Semantic Segmentation

This repository accompanies the paper titled Context-self contrastive pretraining for crop type semantic segmentation published in IEEE Transactions on Geoscience and Remote Sensing. The work presents a novel approach to leveraging supervised contrastive learning for enhancing the performance of semantic segmentation models in identifying crop types from satellite imagery. Performance gains are most significant along parcel/object boudaries which is important for accurate object delineation. No additional data are required to achieve said performance gains.

Getting Started

Environment Setup

Activate the deepsatmodels Python environment to manage dependencies:

conda activate deepsatmodels

Configuration

Dataset Configuration

1. Specify the base directory and paths for training and evaluation datasets in data/datasets.yaml.
2. Configuration files for each experiment are located in configs/. Example configuration files are provided for reference. These files contain default values corresponding to the experimental settings described in the paper.

Model Training and Evaluation

Training Models

To train a model, modify the relevant .yaml configuration file to set the directory for saving models or to load a pre-trained model. The training process can be initiated with the following commands, depending on the model architecture:

• For a randomly initialized UNet3D model:

  python train_and_eval/segmentation_training.py --config_file configs/**/UNet3D.yaml --gpu_ids 0,1
  

• For a randomly initialized UNet2D-CLSTM model:

  python train_and_eval/segmentation_training.py --config_file configs/**/UNet2D_CLSTM.yaml --gpu_ids 0,1
  

• For CSCL-pretrained UNet2D-CLSTM and UNet3Df models, follow the two-step process:

  1. Model pre-training:

     python train_and_eval/segmentation_cscl_training.py --config_file configs/**/MODEL_NAME_CSCL.yaml --gpu_ids 0,1
     

  2. Loading the pre-trained model:

     • Update CHECKPOINT.load_from_checkpoint with the path to the pre-training save directory or specific checkpoint file:

       python train_and_eval/segmentation_training.py --config_file configs/**/MODEL_NAME.yaml --gpu_ids 0,1
       

Replace MODEL_NAME with UNet2D_CLSTM or UNet3Df as appropriate.

Citation

If you use the data or code from this repository in your research, please cite the following paper:

@ARTICLE{9854891,
  author={Tarasiou, Michail and Güler, Riza Alp and Zafeiriou, Stefanos},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Context-Self Contrastive Pretraining for Crop Type Semantic Segmentation}, 
  year={2022},
  volume={60},
  number={},
  pages={1-17},
  doi={10.1109/TGRS.2022.3198187}}