Research on anomaly detection for small-sample optical remote sensing data based on diffusion feature constraints

To address three core challenges of anomaly detection in optical remote sensing imagery under small-sample conditions—complex and heterogeneous data, scarce annotated samples, and insufficient model generalization—this paper proposes a small-sample anomaly detection method based on diffusion feature constraints. Leveraging the robust noise space modeling capability of diffusion models, the method applies multi-scale noise perturbations to input imagery and enforces spatial consistency constraints on features, and detects anomalies via reconstruction error deviation of the encoder-decoder architecture. Experiments on the AID dataset with a convolutional autoencoder (CAE) as the baseline show optimized spatial and spectral entropy, and more complete reconstruction details. Under an extremely small-sample setting with only 1.5%–2.5% anomalous training samples, the method significantly improves sample separability and generalization performance, providing a novel technical pathway for intelligent remote sensing imagery processing.