Abstract: 【Objective】 To establish a convenient and accurate non-destructive detection method for cotton seed germination phenotypes, so as to characterize the salt tolerance of different cotton germplasm at the germination stage. 【Methods】 A synthetic dataset was generated using 150 images of cotton seed germination at different stages and used to train the Mask R-CNN model. Using the trained model, we performed instance segmentation and feature extraction of seed shell and germ in real-world images of 60 cotton germplasm that germinated under 125 mmol/L NaCl treatment, and used them to infer the seed germination rate, germination potential, and germination length, so as to evaluate the salt tolerance of these 60 cotton germplasm in the germination stage. 【Results】 The generated synthetic dataset contained 2,000 images and corresponding mask data. The accuracy of the Mask R-CNN model trained based on this dataset for the segmentation of seed shells and germs in real images was above 95 %, and the phenotypic values that inferred by model were highly consistent with them measured by manual operation (R² > 0.98, P < 0.001), indicating that the phenotypes could be accurately obtained using the model. The cluster analysis of the salt tolerance index for each trait classified the 60 cotton materials into four levels; using the affiliation function method for a comprehensive evaluation of the salt tolerance of the cotton varieties. Kezimian 4 (0.95), MC-30 (0.88), and Lu8zao (0.81) had a larger D-value and indicated high salt tolerance. 【Conclusion】 In this study, we have established a method for phenotyping cotton seed germination traits based on the convolutional neural network model that trained by using synthetic dataset. and using this method, we have identified the seed germination salt tolerance of 60 cotton germplasm in a non-destructive, rapid and accurate manner.