To fully utilize the spatial and spectral information of hyperspectral images and alleviate the problem of insufficient training samples, a semi-supervised hyperspectral anomaly detection model fused with spatial-spectral features is proposed. First, unsupervised clustering is used to automatically construct spatial-spectral background datasets for network training. Then, a dual-path model based on automatic encoder and generative adversarial network is built for learning background spectral features and reconstructing band information respectively. The space branch increases the difference between the background and anomaly using a filter, resulting in spectrum anomaly scores and band outliers. Finally, the spatial-spectral characteristics of anomaly detection are fused. The effectiveness of the proposed method is verified on real hyperspectral images. The experimental results show that the proposed method is superior to the conventional anomaly detection method, with an average detection accuracy of 99.55%.