UniNAD: A Unified Normality-Aware Framework for Anomaly Detection via Self-Supervised Learning

Anomaly detection (AD) in remote sensing aims to identify pixels or regions that deviate from the normal background distribution, without relying on any labeled anomalies. Many unsupervised anomaly detection methods in remote sensing rely on scene-specific background modeling or per-scene network training, which limits the transferability of learned background representations across scenes. Differences in spectral content and spatial structure across sensing modalities further complicate the learning of a shared background representation. To address these limitations, we propose a Unified Normality-aware Anomaly Detection framework (UniNAD) that learns transferable background representations through self-supervised learning on large-scale remote sensing data containing only normal samples. Specifically, we construct a Mamba-driven spectral-spatial sparse encoder (MS3E), which extracts spatial-spectral features with long-range dependencies in the background to learn stable background representations in complex scenes. Furthermore, a semantic prior-guided background prototype set (BPS) extractor is designed. This extractor further integrates high-level semantic priors from the pre-trained DINOv3 model and learns representative prototypes from the fused features to structurally model normal backgrounds. Based on the BPS, we propose a prototype-guided reconstruction (PGR) module. This module uses background prototypes to guide reconstruction toward background-consistent structures and reduce anomaly over-reconstruction. Experiments on hyperspectral, multispectral, and RGB datasets demonstrate that UniNAD achieves the highest mean AUC and mean maximum F1-score among the compared methods across all three modalities. Code and datasets: https://github.com/Zhiru-Yang/UniNAD