Objective: To propose a new solution to overcome the two challenges of data with spikes and small sample sizes in nectarine hyperspectral measurement. Methods: Based on hyperspectral imaging technology, image processing methods were used to identify the area of nectarines in the hyperspectral image, and the spectral reflectance data of the area was calculated to form a hyperspectral curve image. For hyperspectral image data with spikes and noise, compared the effects of several data preprocessing methods, including polynomial smoothing algorithm (SG), multivariate scatter correction algorithm (MSC), standard normal variate algorithm (SNV), first-order derivative operator (D1), and second-order derivative operator (D2) on model prediction accuracy. To address the high-dimensional and small sample size characteristics of the data, the principal component analysis algorithm (PCA) was used for dimensionality reduction, followed by outlier removal using the Mahalanobis distance measure method (MD). Finally, the Kennard-Stone algorithm (KS) was used to divide the data into training and testing sets, and the partial least squares regression (PLSR) model, which performed well in the small sample scenario, was selected for estimation and analysis of nectarine water content. Results: The SG-PCA-MD-KS-PLSR model performed best for estimating nectarine water content when there were spikes and noise in the hyperspectral curve. The coefficient of determination (R²) was 0.928, and the root mean square error (RMSE) was 0.008 4 on the training set. The R² was 0.926, and the RMSE was 0.009 2 on the testing set. In further experiments grading nectarines based on their water content, the model's predictions showed good performance. The accuracy rate of grading was 0.956 for the training set and 0.923 for the testing set. Conclusion: By using hyperspectral imaging technology and establishing the SG-PCA-MD-KS-PLSR model, non-destructive estimation of nectarine water content and grading of nectarine water content can be achieved in scenarios with small hyperspectral sample sizes and noise.