Abstract: In order to solve the problem that many land false alsrms accur in large-scene SAR images ship detection, a method is proposed based on pure background hybrid training to suppress false alarms of large-scene SAR ship detection. The core of this method is to input the image samples without ships (pure background samples) into networks for training, so as to learn the features of pure background samples, and finally realize the false alarm suppression of some strong scattering bright spots on land with high similarity to ships. Due to the lack of pure background image samples in existing public dataset, to verify the effectiveness ofthis method, a pure background hybrid training SAR ship detection dataset that consists of 10 Sentinel-1 large-scene SAR images is also constructed. The experimental results of two kinds of one-stage detectors (Retinanet and SSD) and two kinds of two-stage detectors (Fast R-CNN and Cascade R-CNN) on this dataset show that pure background hybrid training can effectively suppress the false alarms of ship detection in large-scene SAR images.