Rainstorm-induced short-time-series landslides are characterized by strong suddenness, short evolution duration, limited effective samples, and significant rainfall interference. Traditional image segmentation methods are still dependent on manual parameter tuning. Deep learning methods are also difficult to apply stably under limited labeled samples. To address these problems, a fully automatic method for identifying landslide deformation regions is proposed in this study. The method is based on adaptive superpixel segmentation evaluation and fusion. Short-time-series images obtained from a physical simulation experiment of a soil slope under heavy rainfall were used. Geometric distortion correction was first applied to the slope images. A multi-index evaluation system was then constructed. Region consistency, shape regularity, inter-superpixel variance, and segmentation time were included in the system. Principal component analysis was used to determine the optimal superpixel segmentation algorithm and the optimal number of segments. Superpixel fusion was further achieved based on a region adjacency graph and normalized cuts. Target extraction and noise suppression were completed by combining Otsu thresholding and connected-component analysis. A total of nine rainstorm-induced short-time-series landslide images were identified. U-Net and the Otsu method were used as comparative benchmarks. The results show that the proposed method has an area under the curve (AUC) of 0.99. The average accuracy, precision, recall, and F1-score are 99.34 %, 98.77 %, 99.19 %, and 98.97 %, respectively. The overall performance is better than those of U-Net and the Otsu method. The method also shows better boundary recognition capability and robustness on unseen images. Under limited-sample conditions, the proposed method balances recognition accuracy, processing efficiency, and automation. It provides a reference for real-time monitoring and emergency identification of rainfall-induced landslides.