This architecture is a temporal UNet (with ConvLSTMs), featuring an LSTM branch to process climate timeseries and a gating mechanism. It is designed to receive a timeseries of Sentinel-2 images, DEM, and timeseries of climate variables and output a tri-class mask of future MNDWI direction of change.

Developed by: Daniele Rege Cambrin
Model type: ACTU
License: OpenRAIL
Repository: Github
Paper: Arxiv

How to Get Started with the Model

The model is integrated into Transformers, so you can easily load it with the following code:

AutoModel.from_pretrained("DarthReca/actu-direction-classification", trust_remote_code=True, revision=<model_type>)

Load the model with the desired configuration with the revision parameter (the branches of this repo). These configurations are available:

Revision	Backbone	DEM	Climate
main	ConvNeXtV2 Base	No	No
dem-climate	ConvNeXtV2 Base	Yes	Yes

Training Details

The model is pre-trained on Landsat-5 images and fine-tuned on Sentinel-2 of HydroChronos.

Citation

bibtex

@misc{cambrin2025hydrochronosforecastingdecadessurface,
      title={HydroChronos: Forecasting Decades of Surface Water Change}, 
      author={Daniele Rege Cambrin and Eleonora Poeta and Eliana Pastor and Isaac Corley and Tania Cerquitelli and Elena Baralis and Paolo Garza},
      year={2025},
      eprint={2506.14362},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.14362}, 
}

Licensing

The project uses third-party software. For detailed information on the licensing of each component, please see the NOTICE.md file.