Abstract: Abstract: The object detection of Ochotona curzoniae is the basis to count the population and study the population dynamics. In addition, the object detection model based on the deep convolutional neural network requires a lot of sample training. However, the habitat environment of Ochotona curzoniae is harsh and sensitive to change in the external environment, so it is difficult to collect images, resulting insufficient training data on Ochotona curzoniae. The data augmentation method based on Generative Adversarial Network can generate new object images with the same distribution as the original data set, which can effectively solve the problem of insufficient training data for the object detection model. However, when the object image generated by Generative Adversarial Network is fused with the background image, the method of adding pixel by pixel or directly replacing pixel to generate a new image will cause the edge of the fused image to protrude, and when the color difference between the fused target image and the background image is large, the object color of the fused image will be inconsistent with the actual scene. To solve the above problems, this paper proposed an adaptive image fusion data augmentation method based on multi-scale gradients for generative adversarial networks. Firstly, object images are extracted from the training samples and are used to train Multi-Scale Gradients for Generative Adversarial Networks to generate new object images. Secondly, a color histogram is used to select object images and background images with similar colors adaptively. Then, the Poisson fusion method was adopted to fuse the adaptive object image and background image to get a new image, which made the object boundary of the fused image smoother and reduced the color difference between the object and background. Finally, the fusion image was added to the original training set to obtain the augmented training set, and the object detection model was trained. Experimental results of Ochotona curzoniae object detection in natural scenes showed that: the average accuracy of the object detection model trained by the data augmentation method proposed in this paper was 89.3%, which was higher than the average accuracy of the non-data augmentation method. It can improve the detection performance of the object detection model to the Ochotona curzoniae effectively.