CoDeX: Combining Domain Expertise for
Spatial Generalization in Satellite Image Analysis
CoDeX: Combining Domain Expertise for
Spatial Generalization in Satellite Image Analysis
Abhishek Kuriyal1, Elliot Vincent1,2,3, Mathieu Aubry1, Loic Landrieu1,2
1LIGM, ENPC, IP Paris, Univ Gustave Eiffel, CNRS, France
2LASTIG, Univ Gustave Eiffel, IGN-ENSG, 94160, Saint-Mande, France
3Inria, ENS, CNRS, PSL Research University, France
Satellite image analysis struggles with global terrain variation, leading to poor generalization across locations.
We propose a domain-generalization framework that trains one expert model per region and learns their similarities.
A selection module identifies and combines the most relevant experts for each test image.
Our approach outperforms existing methods on four benchmark datasets: DynamicEarthNet, MUDS, OSCD, and FMoW.
We present the different multi-domain training approaches explored in this paper. In (a), we train a single model on all training domains. In (b), we train one model per training domain; all models share the same backbone network, and only see data from their domain. In (c), we add a consistency loss ensuring that the prediction of models associated to similar domains — as defined by a learnable affinity matrix — also produce accurate results.
We present the performance gains achieved by our method over the baseline on all four datasets for three different backbones:
MultiUTAE (our default), ResNet-10, ResNet-18. We observe that our method consistently improves performance across all datasets and all three backbones.
We illustrate for random patches the predictions of our method and our baseline (MultiUTAE). Images from (i-iv) are selected from DynamicEarthNet, (v-vii) from MUDS, and (viii) from OSCD-3ch.
If you find this work useful for your research, please cite:
@article{kuriyal2025combining,
title = {CoDeX: Combining Domain Expertise for Spatial Domain Generalization in Satellite Image Analysis},
author = {Kuriyal, Abhishek and Vincent, Elliot and Aubry, Mathieu and Landrieu, Loic},
booktitle = {EarthVision CVPR Workshop},
year = {2025},
project = {https://github.com/Abhishek19009/CoDEx},
pdf = {}
}
This work was supported by the European Research Council (ERC project DISCOVER, number 101076028) and by ANR project SHARP ANR-23-PEIA-0008 in the context of the PEPR IA. This work was granted access to the HPC resources of IDRIS under the allocation 2024-AD011015600. The authors would like to thank Clémentin Boittiaux, Shiyao Li, Yannis Siglidis, Yohanne Perron, Romain Loiseau, Arijit Ghosh, Eshika Khandelwal, Marta López, Raphaël Baena, Sonat Baltaci, Syrine Kalleli, Guillaume Astruc, Haran Raajesh, and Lucas Ventura for their participation in the annual hackathon held at Imagine and for their valuable contributions to this project.