Abstract: Early non-destructive identification of grafting healing status can improve utilization rate of grafting healing devices．Taking melon grafted seedlings as research object, hyperspectral data of grafted area from the 1st to the 10th day after grafting was obtained．Raw spectral data were preprocessed by different methods, such as the first derivatives（FD）and second derivatives（SD）, standard normal variation（SNV）, detrend processing（Detrend）, smooth 21 points（Smooth 21）, multiple scattering correction（MSC）, and combination of two methods as well．Establishing three classification models: support vector machine（SVM）, decision tree and XGboost, principal component analysis（PCA）, competitive adaptive reweighting sampling（CARS）, genetic algorithms（GA）and continuous projections algorithm（SPA）were used to select characteristic variables corresponding to different classification models．Results showed that FD-GA-XGboost model selected 30 characteristic wavelengths, and prediction accuracy reached 93%．Compared with traditional manual judgment of healing state 10 days after grafting, this method could judge healing state of grafted seedlings more accurately 6 days after grafting, which provided an important reference for production of melon grafted seedlings, and has certain theoretical and practical value．