Hyperspectral imaging technology has been used to establish the prediction model of moisture content in dehydrated scallops, and the model performance is affected by sample division method and modeling method. The method of sample division determines whether the selected sample is representative, and the modeling method determines how to use the sample to build the model, but the internal relationship between the sample division method and the modeling method has been rarely reported. It is important to explore the effects of different sample division methods and modeling methods on the prediction of the moisture content of scallops, and it can also provide reference for the study of spectral modeling of other samples. In this paper, the hyperspectral data of 270 scallops were extracted from spectral images captured by a hyperspectral imaging system in the 380～1 030 nm range. The samples were divided by RS, KS, SPXY and CG. The prediction models were established by PLSR and LS-SVM. The performance indexes of each model were calculated and compared. The results showed that the best sample division method is RS when using PLSR building prediction model (the RPD is 4.079 6) and SPXY is most suitable for LS-SVM model(the RPD is 4.175 6). The advantages and disadvantages of the division of the sample set are related to the modeling method, and the best choice should take modeling method into account. In this commonly used four sample division methods and two modeling methods, SPXY method is used to classify the sample set of moisture content and combine with LS-SVM method to optimize the effect and precision.