Abstract: Considering the radio frequency interference and transient interference in sky-wave over-the-horizon (OTH) radar, the problem of interference detection is converted into range-Doppler (RD) image classification. RD image classifier design and classification performance is discussed. Multi-classifier decision fusion and semi-supervised self-training methods are proposed. Procedure of RD classifier design consisting of RD dataset construction, textual feature extraction, and classification algorithms are introduced. Based on different classification algorithms, various basic classifiers are designed by using the simulated dataset as training set, which obtains accuracy over 95% on strong interference detection but performs poorly on weak interference detection. Hence, ensemble multiple classifiers of different feature views or learning methods are proposed. Besides, algorithms of decision fusion and semi-supervised self-training are proposed. Experiments results on real dataset show that the self-training method based on multi-classifier fusion can improve the recognition accuracy of RD images effectively and increases the detection accuracy of weak interference from below 65% to over 80%.