Convolutional Neural Network (CNN) has a great advantage in data feature extraction, as it can fully acquire data features and has better generalization than traditional models. This study used a hyperspectral prediction method and modeling of Soil Organic Matter (SOM) content based on CNN. Using 320 soil samples from Shangzhuang Experimental Station, Changping District, Beijing, 807 spectral bands within 350~1 700 nm in the visible-near-infrared (VIS-NIR) were extracted, and the spectral data were denoised and transformed by the multivariate scattering correction (MSC) and the first-order differential transform. Successive projection algorithm (SPA) and competitive adaptive reweighted Sampling (CARS) were used to screen the sensitive wavelengths to realize the dimensionality reduction of the spectral data, respectively. To solve the problems of poor generalization of traditional means as well as the complexity and overload of deep CNN networks, based on the CARS and SPA algorithms, a shallow CNN model prediction based on 6 convolutional layers is proposed, and 1D-CNN1 and 1D-CNN2 with different convolutional sizes and number of convolutions are compared to find the optimal network parameters. By comparing the performance of VGG16, Support Vector Regression (SVR), Partial Least Squares Regression (PLSR), and Random Forests (RF) to build a prediction model in the feature wavelength and the full waveform. The optimal model was determined. The results show that compared with the full-spectrum band and SPA filtering algorithms, the model based on CARS filtering feature wavelength modeling performs better, and the number of bands is compressed to 8% of the full-wavelength band, which effectively realizes the dimensionality reduction of the spectral data. Comparing the full-band data, 1D-CNN1 and 1D-CNN2 based on CARS screening wavelengths performed better, with the model predicted R2 improved by 0.028 and 0.018, respectively, and the RMSE reduced by 0.150 and 0.107 g·kg-1, respectively. Overall, the 1D-CNN1 model based on CARS performs the best, with the predicted R2=0.846 and the RMSE decreased by 0.150 g·kg-1, respectively 0.846, and RMSE=3.145 g·kg-1, which reduces the network load while improving the model accuracy, and also proves that small-size convolution outperforms a larger number of large-size convolutions for better acquisition of data features. The SOM content prediction model is established by CARS screening feature wavelengths combined with shallow CNN, which provides a method and reference for establishing a high-precision SOM content prediction model.