Earth surface anomaly detection (ESAD) has become increasingly vital due to intensifying global change and urbanization, leading to more frequent and severe disasters, pollution, and illegal development. Although remote sensing enables wide-range and periodic ESAD, efficiency remains a concern due to lengthy data transmission, dissemination, and processing procedures. Existing methods are primarily task-specific, relying on expert knowledge and human involvement, hindering generalized and automated deployment. In this context, we introduce a novel method aimed at inherently timely on-orbit detection. This approach is not only characterized by its generalizability and automation capabilities, but lightweight parameters that facilitate ground-satellite data transmission for updates. Our approach comprises three main processes: ①employing large vision models as feature extractors to enhance algorithm universality, with automatic feature compression achieved through Gaussian mixture models and Bayesian information criteria, generating lightweight prior knowledge suitable for ground-satellite transmission and on-orbit storage; ②utilizing an efficient dictionary lookup method for rapid inference of surface anomaly scores, making it applicable to satellite-based environments with limited computational resources; ③extracting surface anomaly boundaries based on anomaly scores using prompt words and deep segmentation models. This generates accurate anomaly boundaries with a threshold-insensitive method, reducing human involvement and promoting automation. Experiments demonstrate that the proposed method outperforms traditional approaches, offering better stability and generalization. In experimental cases, the average compression rate of the prior knowledge base was approximately 100 times, significantly improving the ability for on-orbit storage and ground-satellite data transmission updates. Furthermore, the method largely mitigates the issues of low automation and poor noise resistance associated with fixed thresholds for anomaly boundary extraction, achieving automated object-level surface anomaly extraction. Overall, the proposed ESAD method offers advantages such as small data storage, low computational requirements, and high detection accuracy. These features highlight its potential to become a generalized, on-orbit, real-time ESAD for operational deployment.
