Code for the paper On the Generalization of Representation Uncertainty in Earth Observation (ICCV 2025)

📖 Table of Contents

• About
• Installation
• Usage
• Pretrained Uncertainty Checkpoints
• Citation

About

This is the repository for the paper On the Generalization of Representation Uncertainty in Earth Observation. It builds on and extends Pretrained Visual Uncertainties, GitHub repo.

The documentation of this repo will be continuously updated.

Installation

conda create env -f environment.yml

Usage

main.py is the main driver of the project, aggregating the options selected in configs, depending on the task: i.e - Train uncertainties - Inference - Save features (caching)

To initiate an experiment execute python main.py.

The training/inference dataset can be selected by modifying the configs/configs.json "dataset" field.

Training hyperparameters can be modified in configs/train/train_configs.json and inference options in configs/inference/inference_configs.json.

Overal the configuration structure is defined as:

configs
├── configs.json
├── data
│   ├── data_configs.json
│   └── webdataset_configs.json
├── inference
│   └── inference_configs.json
├── stats
│   └── stats.json
└── train
    └── train_configs.json

Pretrained uncertainty checkpoints

Dataset	Bands	ViT-Tiny	ViT-Small	ViT-Base	ViT-Large
BigEarthNet	RGB	Download	Download	Download	Download
BigEarthNet5	RGB	Download	Download	Download	Download
BigEarthNet-SAR	SAR	Download	Download	Download	Download
BigEarthNet-MS	Multispectral	Download	Download	Download	Download
Flair	RGB	Download	Download	Download	Download

ImageNet pretrained models can be found in this repo.

Download example

wget -O bigearthnet_vit_tiny.pth "https://www.dropbox.com/scl/fi/c6j6wa8po22eutyv2w0cd/vit_tiny.pth?rlkey=94u4xcnv1xme2ns93xe7fqor3&st=20uednd6&dl=0"

Citation

If you use our work, please cite:

@misc{kondylatos2025generalizationrepresentationuncertaintyearth,
      title={On the Generalization of Representation Uncertainty in Earth Observation}, 
      author={Spyros Kondylatos and Nikolaos Ioannis Bountos and Dimitrios Michail and Xiao Xiang Zhu and Gustau Camps-Valls and Ioannis Papoutsis},
      year={2025},
      eprint={2503.07082},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.07082}, 
}