Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection - ICCV 2025

This repo contains the code to reproduce the experiments conducted in the paper published at ICCV 2025:
Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection

Implementation of DEFLECT

DEFLECT is based on three elements:

• an Untangled Patch Embedding (UPE): implemented in patch_embed.py and spectral_indices.py (based on Spyndex),
• an Untangled Attention Module (uAtt): implemented in the UntangledAttention class,
• an embedding deflection mechanism: implemented in adapters.py.

Competing methods

Implemented Parameter-Efficient Finetuning methods are:

• LoRA (ICLR 2021),
• Scaled LoRA (CVPR 2024),
• Norm Tuning (ICLR 2024),
• BitFit (ACL 2022).

Reproducing the experiments

Training

Set config, encoder_config, dataset_config, segmentor_config and augmentation_config and start the training process on single gpu:

torchrun --nnodes=1 --nproc_per_node=1 run.py  \
--config configs/run/default.yaml  \
--encoder_config configs/foundation_models/prithvi.yaml  \
--dataset_config configs/datasets/mados.yaml   \
--segmentor_config configs/segmentors/upernet.yaml \
--augmentation_config configs/augmentations/segmentation_default.yaml  \
--num_workers 4 --eval_interval 1
--finetune lora

Evaluation

torchrun --nnodes=1 --nproc_per_node=1 run.py --batch_size 1 --eval_dir work-dir/the-folder-where-your-exp-is-saved

All of these parameters can also be set in the run config file.

To use more gpus or nodes, set --nnodes and --nproc_per_node correspondingly, see: https://pytorch.org/docs/stable/elastic/run.html

To use mixed precision training, specify either --fp16 for float16 and or --bf16 for bfloat16

Citation

@InProceedings{Thoreau_2025_ICCV,
    author    = {Thoreau, Romain and Marsocci, Valerio and Derksen, Dawa},
    title     = {Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {9594-9604}
}