The existing convolutional neural network soil organic matter (SOM) prediction models suffer from low modeling effectiveness and low prediction accuracy under small sample data sets. In order to predict the content of organic matter in the soil more accurately, this paper proposes a hyperspectral prediction model of red soil organic matter with an improved Self Attention Temporal Convolutional Network (SATCN) using 206 collected soil samples as the research object. In this paper, after Savitaky-Golay (SG) smoothing is performed on soil samples, four transformations are performed: first-order differential (1DR), second-order differential (2DR), standard normal variable (SNV) and multivariate scattering correction (MSC). The modeling effects of Long Short-Term Memory (LSTM), Partial Least Squares Regression (PLSR) and Support Vector Machine (SVM) under different spectral preprocessing were compared and analyzed. The results show that the first-order differential preprocessing method of the spectrum after SG processing has the best modeling effect. A shallow network structure is applied in the temporal convolutional network (TCN) architecture, a self-attention layer is added to the TCN residual structure to improve the model feature learning capability, and L2 regularization is added to each convolutional kernel weight to prevent overfitting. First-order differentiation is selected as the spectral preprocessing, and four models of ResNet-13, VGGNet-7, TCN and improved temporal convolutional network (SATCN) are constructed to compare and analyze the modeling effects of the four models, as well as the modeling effects of SATCN models at different network depths. The results show that the shallow SATCN modeling is better than the deep model in the case of first-order differential spectral preprocessing; the self-attention residual structure in the SATCN model not only enhanced the important features of the spectral sequence, but also significantly improved the model feature learning ability and prediction accuracy. Compared with modeling methods such as CNN and TCN, the proposed SATCN model has higher accuracy and excellent model estimation capability with a coefficient of determination (R2) of 0.943, root mean square error (RMSE) of 3.042 g·kg-1, and relative analysis error (RPD) of 4.273 for the validation set. In summary, the best budget SOM content of this paper model is the SATCN prediction model based on the first-order differential spectral preprocessing after SG smoothing, which provides a more accurate prediction of soil organic matter in Guangxi woodlands.