Improvement of rock surface roughness accuracy by combining object space resolution error and 3D point cloud features

To enhance the accuracy of joint roughness coefficient (JRC) estimation in photogrammetry, this study employed a fixed-camera shooting strategy guided by a Structure-from-Motion-based shooting parameter selection algorithm to reconstruct 3D models of rock samples at 16 different shooting distances. The analysis at profile intervals of 0.25 mm, 0.5 mm, and 1 mm revealed a strong correlation between JRC accuracy and three parameters: object space resolution error, spatial distance between point cloud points, and spatial errors of checkpoints on the orientation board. Using these three parameters as input variables and JRC error as the output variable, five machine learning algorithms—Support Vector Regression, Gaussian Process Regression, Multilayer Perceptron, XGBoost, and CatBoost—were employed to predict JRC errors across different shooting distances. The Multilayer Perceptron model performed best at profile intervals of 0.25 mm and 0.5 mm, while XGBoost was optimal at the 1 mm interval. Under the predictions of these models, JRC accuracy improved by an average of 84.7% across the three intervals. Finally, the applicability and limitations of the proposed method were further discussed.