Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models

TL;DR

This work tackles out-of-distribution detection for Earth Observation images, focusing on unsupervised methods that do not require labeled anomalies. It develops ODEED, a reconstruction-based scorer that leverages the deterministic trajectory of the Probability Flow ODE (PF-ODE) in diffusion models to assess image plausibility. Across cloud and SpaceNet 8 flood-domain experiments, ODEED (especially with LPIPS) consistently outperforms baselines in near-OOD scenarios, while diffusion-losses struggle with domain shifts. The approach demonstrates that generative diffusion models can robustly flag rare events and distribution changes in unlabeled EO data, with practical implications for disaster monitoring and data curation.

Abstract

Earth Observation imagery can capture rare and unusual events, such as disasters and major landscape changes, whose visual appearance contrasts with the usual observations. Deep models trained on common remote sensing data will output drastically different features for these out-of-distribution samples, compared to those closer to their training dataset. Detecting them could therefore help anticipate changes in the observations, either geographical or environmental. In this work, we show that the reconstruction error of diffusion models can effectively serve as unsupervised out-of-distribution detectors for remote sensing images, using them as a plausibility score. Moreover, we introduce ODEED, a novel reconstruction-based scorer using the probability-flow ODE of diffusion models. We validate it experimentally on SpaceNet 8 with various scenarios, such as classical OOD detection with geographical shift and near-OOD setups: pre/post-flood and non-flooded/flooded image recognition. We show that our ODEED scorer significantly outperforms other diffusion-based and discriminative baselines on the more challenging near-OOD scenarios of flood image detection, where OOD images are close to the distribution tail. We aim to pave the way towards better use of generative models for anomaly detection in remote sensing.

Figures (10)

• Figure 1: ODEED discriminates between pre and post-event images by making larger reconstruction errors on the latter w.r.t. the LPIPS metric (top). It can also isolate flooded areas from the other post-event images when evaluating the reconstruction similarity with the MSE (bottom).
• Figure 2: Illustration of the one-step denoising and ODEED scorers. (Left) The one-step denoiser samples multiple corrupted versions of the original image $x_{t_0} \sim p_{t_0}(x | x_0)$ thanks to the forward SDE and then evaluates similarity scores on the one-step reconstructions made with the diffusion model $D_\theta$. (Right) The ODEED scorer encodes the initial image into a unique latent $x_{t_0}$ with the PF-ODE estimated with $D_\theta$ and then decodes the latent. For the true PF-ODE, in-distribution samples' reconstruction is perfect.
• Figure 3: The three scenarios derived from the SpaceNet 8 dataset. The first two setups focus on the impact of floodings while the third one centers around geographical domain shift.
• Figure 4: $t_0$vs. AUC$\uparrow$
• Figure 5: $t_0$vs. AUC$\uparrow$