This study aims to improve the conventional empirical rock strength estimation method that is widely adopted in the Australian coal mining industry and provide more accurate rock uniaxial compression strength (UCS) predictions based on artificial neural network (ANN) models. 274 laboratory-measured UCS data (sample dimension: 60 mm (diameter) × 150 mm (height)) were collected from two longwall coal mines in Australia, along with the four geophysical logs (sonic, gamma, neutron, and porosity logs) and rock density of the tested core specimens. A two-layer ANN model was developed from these data based on the Levenberg–Marquardt algorithm as a base model. As compared to conventional sonic-UCS fitting equations, the mean percentage error, root mean squared error and maximum absolute error of the proposed ANN model were reduced from 34.27%, 15.23 MPa and 70.66 MPa to 20.67%, 11.02 MPa and 47.66 MPa, respectively. Further investigations were conducted by splitting the dataset based on rock lithology and mine locations, and separate ANN models were developed based on the divided subsets. Both lithology-specific and site-specific models exhibited improved estimation accuracy in comparison to the base model, indicating lithology and local geological conditions could considerably affect the estimation accuracy. Among the three types of models, the lithology-specific models yielded the most accurate predictions over the 274 data (MPE: 17.55%; RMSE: 8.84 MPa), followed by the site-specific models and the base model in descending order. Overall, all three models significantly outperformed the current empirical methods adopted in the studied mine sites. The outcomes of this study could provide more accurate UCS predictions for further geotechnical analysis in underground constructions, such as tunnel stability analysis and estimating in-situ stresses based on borehole breakout.