The convolutional neural network (CNN) has been widely used in various chemometric tasks in the past few years. However, learning long-range correlations from spectra using the CNN remains challenging, because most CNN architectures utilized in previous studies are quite shallow to avoid overfitting. In this paper, we present an atrous convolutional network (ACPnet) for learning long-range spectral correlation in quantitative spectrometric analysis. Paralleled convolution branches with different atrous rates are assembled to determine the best trade-off between short-range and long-range information. Three data sets, viz. tablets (Raman), soil (NIR), and wines (NMR), are evaluated to demonstrate the versatility of the proposed network. The overall results indicate that the ACPnet achieves better regression accuracies for all three data sets than those of partial least squares regression (PLS), least square support vector machine (LS-SVM), a regular CNN, and an atrous CNN in a cascaded pattern (ACCnet). Furthermore, the features extracted by the ACPnet are fed into different regressors to evaluate the proposed network as a supervised feature extractor. The results of the extraction–regression model show that ACPnet gives better feature-extraction performance than that of a conventional CNN on the three data sets.