Abstract: The traditional partial least squares method and support vector machine regression method were often difficult to obtain high accuracy and further optimization in predicting the element content of the ground standard sample of the rover corresponding to the spectrum. To solve the above problems, the three-channel folding of high-dimensional spectral information was carried out to eliminate its matrix effect in the research, and introduced the Residual Network structure (ResNet), which was good in the field of computer vision, to extract the spectral features and predict the corresponding principal component content. In this paper, the full connection layer in ResNET network structure was removed to prevent the sudden increase of model parameters, and the last Softmax classification sublayer of the network was changed into a linear rectification layer for prediction. At the same time, exponential learning rate attenuation and Dropout mechanism were added to make the model prediction results have higher accuracy and generalization ability. Compared with linear support vector machine regression (LinearSVR) and depth separable convolution network Xception, the prediction root mean square error of each main element content of the model decreases by 30% and 17% on average, respectively. The experimental results show that the regression model established by ResNet network shows good prediction characteristics when LIBS technology is used for principal element quantitative analysis of ChemCam spectral data.