FlowEO: Generative Unsupervised Domain Adaptation for Earth Observation

TL;DR

FlowEO tackles the challenge of distribution shifts in Earth Observation by performing unsupervised domain adaptation directly in image space using latent flow matching. It learns a semantically preserving transfer between source and target image distributions in a latent space, guided by a data-dependent coupling and an ODE-based inference pipeline, making downstream tasks agnostic to the adaptation method. Empirically, FlowEO consistently outperforms standard image-translation baselines in segmentation and classification across multiple EO datasets and scenarios, while delivering superior perceptual image quality without adversarial losses. The approach enables robust cross-domain EO analysis without retraining downstream models and points to future extensions using metadata-aware couplings for unpaired scenarios.

Abstract

The increasing availability of Earth observation data offers unprecedented opportunities for large-scale environmental monitoring and analysis. However, these datasets are inherently heterogeneous, stemming from diverse sensors, geographical regions, acquisition times, and atmospheric conditions. Distribution shifts between training and deployment domains severely limit the generalization of pretrained remote sensing models, making unsupervised domain adaptation (UDA) crucial for real-world applications. We introduce FlowEO, a novel framework that leverages generative models for image-space UDA in Earth observation. We leverage flow matching to learn a semantically preserving mapping that transports from the source to the target image distribution. This allows us to tackle challenging domain adaptation configurations for classification and semantic segmentation of Earth observation images. We conduct extensive experiments across four datasets covering adaptation scenarios such as SAR to optical translation and temporal and semantic shifts caused by natural disasters. Experimental results demonstrate that FlowEO outperforms existing image translation approaches for domain adaptation while achieving on-par or better perceptual image quality, highlighting the potential of flow-matching-based UDA for remote sensing.

Figures (10)

• Figure 1: FlowEO generates realistic and semantically consistent outputs on various challenging image translation tasks, such as pre- to post-disaster domain adaptation and SAR-to-Optical translation.
• Figure 2: FlowEO learns a latent flow between the source and target distributions in four stages: 1) the training image pairs are sampled from the coupling $p(x_0, x_1)$, 2) images are encoded in SD3 latent space, 3) we interpolate between the latent codes $z_0$ and $z_1$ to compute $z_t$ for $t\sim \mathcal{U}(0, 1)$, 4) we train the U-Net backbone $v_\theta$ on a simple regression loss to match the conditional velocity $u_t(z_t\mid z_0, z_1)$.
• Figure 3: FlowEO offers domain adaptation in image-space, making the adaptation independent of the downstream task and predictive model used. At inference time, we adapt the image $x_0$ into a synthetic image $\hat{x}_1$ by integrating the flow with an ODE solver and the learned velocity $v_\theta$. Then, any predictive model $S_1/C_1$ can directly perform downstream tasks on the transferred images, without fine-tuning.
• Figure 4: Weakly-aligned image pairs from the SpaceNet 8 dataset, affected by cloud coverage and natural disasters. Each column: top=post-flooding imagery; bottom=pre-event imagery.
• Figure 5: Qualitative comparison of domain adaptation methods on segmentation datasets. The first column represents the input image $x_0$, the second and third depict the weakly or strongly aligned $x_1$, and the others display the images generated by the different methods. Below each image, we provide the corresponding prediction from the segmentation model $S_1$ or the true segmentation mask $y_1$ for the reference image (third column). FlowEO outperforms other methods in both semantic preservation and image quality.