•DCNNs were used for regression and classification predictions of soil properties.•DCNNs outperformed single-task shallow CNN and traditional approaches.•DCNNs do not need spectral dimension reduction for modeling soil spectral data.•The position of feature wavelengths differed among the soil properties. Soil, as a non-renewable resource, should be monitored continuously to prevent its degradation and promote sustainable agriculture. Soil spectroscopy in the visible-near infrared range is a fast and cost-effective analytical technique to predict soil properties. Although traditional machine learning methods are widely used for modeling soil spectral data, large spectral datasets may require better analytical methods for big data. Here, we explored the modeling potential of deep convolutional neural networks (DCNNs) for soil properties based on a large soil spectral library. The European topsoil dataset provided by the Land Use/Cover Area frame Survey (LUCAS) was used for DCNN modeling with the original absorbance spectra. Two single-task 16-layer DCNN models (LucasResNet-16 and LucasVGGNet-16) were used to make regression predictions of seven soil properties and classification predictions of soil texture. The effects of data pre-processing on single-task and multi-task DCNN modeling were assessed. The SHapley Additive exPlanations method was used to interpret the output of a DCNN model (LucasResNet-16). The DCNN models produced accurate predictions for most soil properties, and were superior to a single-task shallow convolutional neural network and traditional machine learning methods. Spectral transformation was effective for predicting some soil properties, while spectral downsampling led to a reduction in the modeling accuracy. The performance of a multi-task DCNN model built on the basis of LucasResNet-16 was improved compared with the performance of the single-task model. Soil organic carbon content, nitrogen content, cation exchange capacity, pH, and calcium carbonate content were well predicted, with the root mean squared error of 19.130 g∙kg−1, 0.971 g∙kg−1, 6.614 cmol(+)∙kg−1, 0.326, and 24.526 g∙kg−1, respectively. The overall classification accuracy of soil texture was 0.749 (four groups) and 0.566 (12 levels). The position of feature wavelengths differed among the soil properties, for which multiple characteristic peaks were common. This study fully demonstrates the modeling potential of deep learning with soil ultraspectral data, which could enhance precision agriculture.