Towards Onboard Continuous Change Detection for Floods

This repository contains the official implementation of the HiT (History-Injection Transformer) mechanism, as described in the paper: "Towards Onboard Continuous Change Detection for Floods."

HiT maintains historical context from previous observations while reducing data storage by over 99% of original image size (compared to bi-temporal baseline). Testing on the STTORM-CD-Floods dataset confirms that the HiT mechanism within the PrithVi-tiny foundation model maintains detection accuracy compared to the bi-temporal baseline.

Model F1 Precission Recall Parameters Input Size
Baseline 0.41 Β± 0.06 0.73 Β± 0.05 0.29 Β± 0.05 8.5M 2Γ—
HiT-PrithVi 0.38 Β± 0.08 0.70 Β± 0.03 0.27 Β± 0.08 7.8M 1.004Γ—
ContUrbanCD 0.46 Β± 0.26 0.82 Β± 0.06 0.35 Β± 0.25 25M nΓ—

πŸ“Š Datasets

Installation

git clone https://github.com/zaitra/HiT-change-detection.git
cd Hit-change-detection
pip install .

πŸ“œ Citation

If you use this work, please cite the following paper:

@misc{kyselica2026towards,
      title={Towards Onboard Continuous Change Detection for Floods}, 
      author={Daniel Kyselica and JonÑő Herec and Oliver Kutis and Rado PitoňÑk},
      year={2026},
      eprint={2601.13751},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.13751}, 
}

Maintained by Zaitra.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Paper for ZAITRA/HiT-Prithvi