Spectroscopic modeling of soil element content based on machine learning and deep learning is a research hotspot in soil chemical composition detection. In order to improve the accuracy of soil element content spectroscopy modeling based on convolutional neural network, a particle swarm optimization optimized convolutional neural network soil nitrogen element content spectral analysis model is proposed. Smoothing and standard normal transformation of soil samples was used to reduce the impact of noise on modeling. A convolutional neural network structure suitable for regression was designed. The hyperparameters of convolutional neural network such as kernel parameters, learning rate, number of iterations were optimized by using particle swarm optimization. The wavelength range of the visible/near infrared spectrum is 225~975nm, by analyzing and modeling the nitrogen content of 177 groups of soil samples in Qingdao area, the model’s results show that, compared with PLS, CNN and other modeling methods, the PSO-CNN model proposed in this paper has a higher prediction accuracy. The coefficient of determination of the test set is 0.9707, the root mean square error is 0.8818, and the ratio of performance to standard deviation is 5.88.