Gradation regression prediction for engineering based on multiscale rockfill instance segmentation

Abstract

The gradation of rockfill material is crucial for the dam stability. The inefficiently of manual screening poses a challenge to the rapid and accurate spatial grading detection in construction projects. This paper proposes a machine learning method, which realizes large-area multiscale rockfill gradation prediction through rapid analysis UAVs imagery. The novel Rock-net model, incorporating advanced feature extraction technologies, achieves efficient segmentation dense rockfills. The morphological feature extraction and classification statistics of rockfills in high-resolution images have been achieved utilizing SAHI. The CPO-XGB regression model adeptly delineates the nonlinear relationships between feature parameters and spatial gradations, demonstrating superior causal prediction capabilities. The experimental results demonstrate that the Rock-net achieves a Box_AP of 90.8%, Seg_AP of 90.5%, and FPS of 12.8 it/s, while CPO-XGB attains an R2 of 0.977 and MSE of 0.00032, significantly better than similar models. The principle of similar fractals quantifies the gradation characterization, making it applicable for gradation detection across different particle size ranges. Both laboratory tests and field applications validate the effectiveness and exceptional generalization potential of this approach.