Soil total nitrogen is an important nutrient index of soil. The soil of pear orchard of Jiangsu Academy of Agricultural Sciences was taken as the research object, the soil spectral reflectance data were obtained by hyperspectral imaging technology, the shuffled frog leaping algorithm and competitive adaptive reweighted sampling in total nitrogen content in orchards was studied and constructed based on hyperspectral data, which provided a method for accurately detecting soil total nitrogen content. The competitive adaptive reweighted sampling were introduced for spectral feature extraction, and the partial least squares regression, support vector regression, random forest and convolutional neural network models were used to estimate the total nitrogen content of the soil by using the full band and the characteristic band, respectively. The results showed that after the original spectrum was processed by a variety of preprocessing methods, it was found that the total nitrogen prediction model constructed by SG convolution smoothing combined with standard normal transform pretreatment had the best performance. Based on the shuffled frog leaping algorithm, totally ten key bands were extracted, accounting for 4.08% of the total number of bands, which effectively reduced the data dimension. The convolutional neural network model constructed based on the shuffled frog leaping algorithm to extract feature bands performed well, and the coefficient of determination of the model test set was 0.95, the root mean square error was 0.21g/kg, and the relative analysis error was 3.97. The results showed that the shuffled frog leaping-algorithm can efficiently extract the feature bands, reduce the data dimension, and improve the estimation accuracy of soil total nitrogen content, which provided a reference for the accurate estimation of soil total nitrogen content in orchards.