ADMamba: A Spatio-Temporal Mamba-Inspired Model for Anomaly Detection in Satellite Videos

View at publisher Summarise with VIRAT
Authors: Zhi He, Jun Li, Xueqiang Zhao, Chengle Zhou, X. Mei
Cited by: 0
References: 90

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Anomaly detection (AD) has become an increasingly significant research topic for interpreting and understanding remote sensing images. However, most existing studies focus on hyperspectral and high-resolution images, leaving satellite videos relatively unexplored. Compared to these images, AD in satellite videos presents greater challenges due to lower spatial-spectral resolution and the need to process a larger number of frames. In this paper, we propose a spatio-temporal Mamba-inspired model (ADMamba) for AD of satellite videos. The proposed ADMamba consists of two main modules: a clustering-based coarse detection module and a Mamba-based fine detection module. First, note that the anomalies are small and distinct from their surroundings but should be detected without any prior knowledge, a clustering-based module is proposed to detect the pseudo coarse background and anomaly samples. Second, the Mamba-based module is composed of three channels formed from one-dimensional (1D), two-dimensional (2D), and three-dimensional (3D) U-Net architectures. This design fully exploits spatial and temporal information, enhancing the feature representation capability of satellite videos. Contrastive losses are adopted to maximize the reconstruction error of pseudo anomalies while minimizing the reconstruction error of pseudo backgrounds during the training process. In the testing process, the AD results are obtained by calculating the mean reconstruction error across the three channels. Experiments conducted on three Jilin-1 and two OVS-1 satellite videos demonstrate that the proposed method outperforms the state-of-the-art AD methods. The code for this paper is available at https://github.com/hzhdhz/ADMamba.

Fields of Study & Subjects

Remote-Sensing Image Classification Anomaly Detection Techniques and Applications Satellite Image Processing and Photogrammetry

Publication details

Journal / source
IEEE Transactions on Geoscience and Remote Sensing
Volume
64
Pages
1–21
Publication date
1 January 2026
Language
EN