<p>Corrosion is inevitable for steels served in aqueous environments, and corrosion wastage prediction must be carried out, so as to provide guidance for entire life cycle design of new built steel structures, as well as residual performance assessment of ageing structures. To this end, corrosion data of stainless steel, low-alloy steel, and carbon steel exposed to eight aqueous environments of China is summarized first in this paper, and then the environmental factors including water conductivity and tidal range are set as independent variables to conduct corrosion rate prediction using four machine learning (ML) algorithms, i.e., GRNN, SVM, RF, and XGBoost. Subsequently, the constructed ML prediction models are interpreted by SHAP, to analyze the impact weight of each environmental factor on corrosion behavior. Finally, the environmental factors are selected through the values of SHAP, and the trivial factors that affect corrosion are phased out, so as to re-predict the results via the reserved key parameters. The advantages of prediction through reduced parameters over that through complete parameters are then demonstrated. The results show that the <i>R</i>² values of corrosion prediction under all cases are greater than 0.9, indicating all the four ML algorithms can yield acceptable prediction results for corrosion wastage of steels exposed to aqueous environments. Among all the considered 15 environmental factors, six factors (water conductivity, tidal range, content of sulfate, content of chloride, pH, and oxidation reduction potential) are main factors that affect the corrosion behavior of steels exposed to aqueous environments, while the other nine factors have marginal influence on corrosion rate. Interpretable ML can yield better prediction results in the case of reducing input parameters and decreasing computational cost.</p>