Abstract:

Addressing the challenges of complex data sources and insufficient model generalization ability in small-sample optical remote sensing anomaly detection, this paper proposes a diffusion feature-constrained small-sample optical remote sensing anomaly detection method. By introducing the noise space modeling capability of the diffusion model, this method enhances the stability and capability of feature learning, and achieves anomaly detection in small-sample scenarios based on the deviation degree of reconstruction error. Using the Wuhan University AID dataset as experimental data, a comparative experiment was conducted between the proposed method and the convolutional autoencoder benchmark method. The experimental results show that the proposed method reduces the spatial entropy mean from 3.65 to 3.51 and the spectral entropy mean from 5.77 to 5.62, with significant improvements in quantitative indicators. Furthermore, the reconstruction results are visually more complete and less noisy. The anomaly detection experiment simulates common image anomalies such as unclear textures and striping noise in national standards. In small-sample scenarios where the proportion of abnormal samples in the training set is 1.5%to 2.5%, subjective visual evaluation and quantitative index analysis indicate that the negative samples are well separable in most land types. This paper verifies the effectiveness of diffusion feature constraints in small-sample anomaly detection, providing ideas for optical remote sensing quality assessment.

Key words: reconstruction error, small sample, anomaly detection, diffusion model, convolutional auto-encoder