Predicting butterfly species presence from satellite imagery using soft contrastive regularisation

This repository contains all code of our 2025 CVPR FGVC paper, including:

• PECL (Paired Embeddings Contrastive Loss) implementation in scripts/paired_embeddings_models.py.
• Torch dataloader for the S2BMS dataset in scripts/DataSetImagePresence.py
• Resnet-based model to predict species presence vectors from satellite images, using PECL.

Installation:

• Use conda to install packages using pecl.yml or pip install from requirements.txt.
• Add your user profile data paths in content/data_paths_pecl.json. (This step is not needed when just experimenting with the code and the example data provided in the repo).

Getting started:

• A sample data set (of 16 locations) is provided in tests/data_tests/.
• Go to notebooks/Getting started.ipynb to see examples of how to load the data and model.

Data:

• The full S2-BMS data set is available on Zenodo.
• Our Torch dataloader is available in scripts/DataSetImagePresence.py.

PECL implementation

• For details please see our paper.
• PyTorch implementation can be found in scripts/paired_embeddings_models.py (ImageEncoder.pecl_loss()).

Results

• The training scripts used for the paper are scripts/train.py and scripts/train_randomsearch.py.
• The figures and tables in the paper were created in notebooks/Results figs and tables.ipynb.

Please cite our paper if you use this method or data in a publication - thank you!!