MIRANDA is a domain adaptation method designed for species-level phenology forecasting. The backbone architecture is PhenoFormer, which is a transformer model designed to predict the phenophases of multiple species at the same time from an input meteorological time seres. The proposed MIRANDA addresses domain shifts in phenology modelling via two key components: rank-based adversarial training on intermediate features (green parts) and hybrid layer normalization (blue parts).

We use the same dataset as in the PhenoFormer paper. You can retrieve it from the Zenodo archive. This dataset contains two subfolders: one version of the dataset formatted for R scripts and the other one for python scripts. Please use the learning-models-data subfolder for all python scripts.

git clone [email protected]:SherryJYC/MIRANDA.git
cd MIRANDA
conda create --name miranda python==3.12
conda activate miranda
pip install -r requirements.txt

For the deep learning scripts, we recommend using a machine with GPU to have reasonable training times. Our models are still quite small (for deep learning standards) so a small GPU of even 4 or 8GB VRAM would do.

Main script to run the configurations of MIRANDA:

• run-phenoformer-multispecies-spring.py to train the multi-species variants for spring phenology.

To run the main script:

1. Complete the data_folder field with the path to the learning-models-data dataset folder on your machine.
2. Complete the save_dir field with the path to the folder where to write the results.
3. Activate the proper python environement and run the script.

Different models used in the paper are defined in model_configs in configs/RUN_CONFIGS.py, you can choose the model name and put in dict_model_to_do_list in run-phenoformer-multispecies-spring.py.

• the proposed method: MIRANDA
• baselines: dann, adda, adaBN, CORAL
• ablations: dann_shallow (dann on mid-feature), dann_shallow_rank_cos (dann on mid-feature + rank loss), dann_shallow_daln_nf (dann on mid-feature + hybrid Norm)

There are three datasets used in the paper, and you can choose which dataset to use inside run-phenoformer-multispecies-spring.py.

• structured_temporal: Chronological in the paper.
• hotyear_temporal: Annual temperature in the paper.
• highelevation_spatial: Elevation in the paper.

To cite this work please use:

@article{jiang2026miranda,
          title={MIRANDA: MId-feature RANk-adversarial Domain Adaptation toward climate change-robust ecological forecasting with deep learning},
          author={Jiang, Yuchang and Wegner, Jan Dirk and Garnot, Vivien Sainte Fare},
          journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
          year={2026},
        }

• Data source: Federal Office of Meteorology and Climatology (MeteoSwiss)
• Meteorological data processing: Swiss Federal Institute for Forest, Snow and Landscape Research (WSL)
• Codes are largely adopted from the PhenoFormer official repository.