Smart Prediction of Carbon Emissions in Bridge Construction: An Empirical Study Using Multi-Dimensional Machine Learning

Abstract

Transportation infrastructure construction generates substantial greenhouse gas emissions, yet construction-phase emissions from small- and medium-span concrete bridges remain difficult to quantify at the design stage. This study develops a data-driven framework for carbon accounting and prediction using 58 urban concrete bridges in Guangdong Province, China, comprising 23 hollow slab girder bridges and 35 concrete box girder bridges. Construction-phase emissions were quantified using process-based life cycle assessment and inventory analysis. Three prediction approaches—Kriging, Kriging–support vector machine, and sequential sampling–ISC–Kriging—were evaluated across low-, medium-, and high-dimensional variable spaces. Material production dominated the carbon footprint, accounting for 84.7% of emissions from hollow slab bridges and 82.0% from box girder bridges. Average unit-volume emissions were 499.81 and 686.78 kg CO2e/m3, respectively. The sequential sampling–ISC–Kriging model achieved the best performance in the high-dimensional space (R2 = 0.920), and its error for the independent case bridge was 5.65%. These findings show that integrating structural, construction, and transportation variables can support early-stage carbon estimation and comparison of low-carbon bridge alternatives.