Soil properties are vital to profiling and utilizing soil resources. Conventional approaches to measurements of soil properties often involve costly, environmental-unfriendly, and time-consuming laboratory procedures. Conversely, machine learning (ML) and deep learning (DL) are gaining traction in giving rapid, non-destructive, and cost-saving alternatives to predictions of soil properties. These ML/DL models are convenient and fast because they utilize spectral data, such as visible and near-infrared (Vis-NIR) spectra, that can be easily collected using proximal sensors for their training and prediction purposes. However, existing ML/DL approaches to this problem pose several limitations, such as having small sample sizes, needing to divide the sample data into local areas to increase accuracy, and having relatively low accuracy. Therefore, this work experiments various ML/DL methods that leverage Vis-NIR spectra collected from a rather large number of soil samples distributed all over the world to predict pH H2O and pHKCl. We then propose a DL method, called RDNet, that outperforms the other existing approaches. We also utilize visualizations to verify if the proposed model learns legitimate information from the training data.