Abstract: The longissimus dorsi in pigs was prepared into the meat samples with water injection ratio of 0%, 2%, 6%, 10% and 14% by direct injection. The NMR signals generated by the meat samples were measured by LF-NMR technique. The transverse relaxation spectrum parameters obtained by inversion were taken as independent variables, combined with discriminant analysis and partial least square regression (PLSR) analysis, the models for detecting water-injected meat were developed, and multiple possibilities for PLSR modeling were attempted. It was showed that the developed discriminant model was validated with calibration set and leave-one-out cross, the total accuracy of the classification of water-injected meat was 89.4% and 88.2%, respectively. By combining the lever value with the student residual, the abnormal data was distinguished and deleted, and the prediction performance of the PLSR model established by using the variable projection importance analysis method to select out six transverse relaxation spectrum parameters as independent variables was optimal, the determination coefficient (R_c²) and standard error (SEV) from calibration set were 0.9603 and 1.0033%, the determination coefficient(R_cv²) and standard error (SECV) from cross validation were 0.9508 and 1.1169%, the determination coefficient(R_p²) and standard error (SEP) from prediction set were 0.9518 and 1.1280%, respectively. The best estimate of the confidence interval capable of predicting the percentage of water injection in unknown samples was about 2.256% at 95% confidence probability. The results showed that the DA model and PLSR model based on LF-NMR transverse relaxation spectrum could be used for qualitative and quantitative detection of water-injected meat.