CoDEx: Combining Domain Expertise for Spatial Generalization in Satellite Image Analysis

TL;DR

CoDEx addresses spatial domain shifts in satellite imagery by training $D$ domain-specific experts with a shared backbone and enforcing cross-domain consistency via a learnable affinity matrix $a obracket obreakin obreakin \\mathbb{R}^{D\times D}$ whose row-wise softmax excludes the diagonal. A separate domain-expert selection head predicts aggregation weights, enabling test-time mixtures without target-domain fine-tuning and using losses $ ext{L}_{ ext{domain}}$, $ ext{L}_{ ext{con}}$, $ ext{L}_{ ext{acc}}$, and $ ext{L}_{ ext{mix}}$ to guide learning. The framework is evaluated on DynamicEarthNet, MUDS, OSCD, and FMoW across segmentation, change detection, and classification, consistently outperforming domain-adaptation and domain-generalization baselines. By learning domain similarity and aggregating expert predictions, CoDEx achieves robust spatial generalization with modest test-time overhead and publicly releases its code for reproducibility and practical deployment.

Abstract

Global variations in terrain appearance raise a major challenge for satellite image analysis, leading to poor model performance when training on locations that differ from those encountered at test time. This remains true even with recent large global datasets. To address this challenge, we propose a novel domain-generalization framework for satellite images. Instead of trying to learn a single generalizable model, we train one expert model per training domain, while learning experts' similarity and encouraging similar experts to be consistent. A model selection module then identifies the most suitable experts for a given test sample and aggregates their predictions. Experiments on four datasets (DynamicEarthNet, MUDS, OSCD, and FMoW) demonstrate consistent gains over existing domain generalization and adaptation methods. Our code is publicly available at https://github.com/Abhishek19009/CoDEx.

Figures (5)

• Figure 1: Combing Domain Experts. The appearance of satellite images can vary significantly across regions, even when they contain the same semantic features (here, coastlines). We train experts for each discrete location in the train set, and learn to aggregate the most relevant ones when confronted with an unknown domain.
• Figure 2: Multi-Domain Training. We present the different multi-domain training approaches explored in this paper. In \ref{['fig:over:baseline']}, we train a single model on all training domains. In \ref{['fig:over:multi']}, we train one model per training domain; all models share the same backbone network, and only see data from their domain. In \ref{['fig:over:multiours']}, 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. indicates vector multiplication and a module with tunable parameters.
• Figure 6: Domain Expert Selection. We freeze the models trained previously, and train a domain expert selection model to select the most relevant models for a given input sample from an unseen domain. : vector multiplication, : frozen layers, : module with tunable parameters.
• Figure 7: Qualitative Segmentations. We illustrate for random patches the predictions of our method and our baseline. Images from (i-iv) are selected from DynamicEarthNet, (v-vii) from MUDS, and (viii) from OSCD-3ch.
• Figure 8: Impact of Backbone. We report the performance of the baseline and our approach for all four datasets and three backbones networks.