<p>The estimation of the joint roughness coefficient (JRC) is a critical aspect of rock mass mechanical property evaluation. However, existing quantitative evaluation models for JRC often suffer from construction difficulties and low prediction accuracy. To address this issue, a dataset comprising 112 joint profiles was collected. The dataset consists of 112 samples, each described by eight extracted features (i.e., an 112 × 8 feature matrix). The sparrow search algorithm (SSA) was employed to optimize the hyperparameters of a multilayer perceptron (MLP) model, and a novel hybrid machine learning model combining SSA and MLP was developed. SSA is a population-based swarm intelligence optimizer that iteratively improves a set of candidate solutions by balancing global exploration and local refinement. The predictive performance of the proposed SSA-MLP model was compared with that of a conventional MLP model and other empirical models. The results indicate that the proposed SSA-MLP model significantly outperforms the MLP model, achieving a coefficient of determination (R²) of 0.9574 and a root mean square error (RMSE) of 0.9480. Furthermore, feature importance analysis revealed that standard deviation of inclination angle is the most influential factor affecting JRC estimation. This study provides new insights into the application of machine learning in JRC prediction and lays the foundation for achieving high-precision JRC estimation.</p>