In recent years, the rapid development of modern spectral detection technology is closely related to deep learning. As an end-to-end model, the deep neural network can get more information from the spectra, thus improving the robustness of the model. A one-dimensional residual neural network (1D-AD-ResNet-18) model based on a convolutional block attention module was proposed to explore the feasibility of qualitative prediction of mango species by near-infrared spectroscopy combined with deep learning. Firstly, to reduce the interference of redundant information in the spectra, the CBAM convolution attention module is added to the traditional one-dimensional residual neural network, which can focus on the local useful information of the spectra. Secondly, to avoid the disappearance of gradient and the occurrence of overfitting, ResNet-18 is used to solve the problem of network “degradation”. For 186 mango samples, 70% of the samples were trained, and 30% were tested. Accuracy, Precision, Recall, F1-score, Macro-average, and weighted average were used as evaluation indexes of the model. Three comparison models were established, including traditional one-dimensional ResNet-18, SNV-SVM, and PCA-KNN. Compared with the above three methods, the established 1D-AD-ResNet-18 model obtained the optimal prediction results, and the accuracy of the four qualitative analysis models was 96.42%，80.35%，76.78% and 67.85%. The experimental results show that the 1D-AD-ResNet-18 model can accurately identify and classify mango species, which provides a new idea for the qualitative analysis of mango species by NIR spectroscopy.