Predicting Aggregate Gradation Based on Pavement Surface Image Features Using Random Forest and Multilayer Perceptron Integrated Model

Aggregate gradation significantly impacts the road performance of asphalt pavement. However, current detection methods are often inefficient, destroying the pavement structure, and do not allow for quick real-time detection. The advancement of machine learning presents an effective solution for the rapid detection of aggregate gradation, which can be achieved by analyzing pavement surface images and leveraging extensive feature data to predict aggregate gradation. This article discusses the completed asphalt pavement surface, the use of CMOS cameras, LED light panels, and shade boxes to build a field image acquisition device to obtain pavement images and standardize the image acquisition standards; it also addresses the problems of low segmentation accuracy and the time-consuming nature of current pavement image processing. We thus propose a method for segmenting adherent aggregate particles using an integrated artificial intelligence model and extracted 47 image features, wherein a total of 1,440 image samples were acquired using pavement images of 18 aggregate gradations; further, a random forest and multilayer perceptron integrated model (RF-MLP) was developed to predict the aggregate gradation of asphalt pavements. The results indicate that the RF-MLP integrated model for detecting asphalt pavement aggregate gradation outperforms single models. The goodness-of-fit (R2) for the test samples reaches an impressive 0.893, while the average error in the sieve passing rate is only 1.419%, demonstrating a high level of accuracy in the prediction results. Moreover, the rapid and nondestructive detection of pavement aggregate gradation can be achieved by capturing images of the pavement surface after compaction.