基于随机森林算法的CPTu土类识别模型研究及其在不同区域的应用

Research on CPTu-based soil classification model using random forest algorithm and its application in different regions

探讨基于跨地区的“CPTu+钻孔”数据库建立多地区广泛适用的土类识别模型的可行性,阐明在砾石、砂土、粉土、黏土四大分类情况下,能够适用于多个不同地区的土类识别模型。基于跨新西兰、奥地利、德国3个地区的“CPTu+钻孔”数据库,以CPTu测试数据的8个统计特征为输入,以砾石、砂土、粉土、黏土4种土类为输出,应用随机森林算法建立分类模型,同时探讨RF、SVM、BPANN、KNN 4种机器学习算法对于该土类识别问题的性能差异。研究结果表明:模型在3个地区均具有良好的泛化性能,与工程中广泛接受的SBTn图表法相比,土类预测精度显著提升。结合该模型和相应的土层界面确定方法,可应用于完整的CPTu测试曲线以重构测点处的土层分布。重构土层分布和钻孔土层分布具有很好的一致性,一致性程度达95%左右。在4种机器学习算法中,RF算法具备最优的性能,能有效解决不平衡分类问题。

The feasibility of building a multi-regional soil classification model based on the cross-regional"CPTu+borehole"database was investigated,and it illustrates that a single soil classification model can be suitable for multiple regions with four major classifications:gravel,sand,silt and clay.A"CPTu&borehole"database from New Zealand,Austria and Germany was established,and a soil classification machine learning model was developed based on random forest algorithm using the eight statistical characteristics of CPTu data as inputs and four soil classes,including gravel,sand,silt,and clay,as outputs.Furtherly,the performance of four kinds of machine learning algorithms,namely RF,SVM,BPANN and KNN,were discussed in detail for CPTu-based soil classification.The results show that the soil classification model has good generalization performance in three regions,i.e.,New Zealand,Austria and Germany,and exhibits remarkable better performance than SBTn method.Combined with an appropriate soil boundary determination method,the model can successfully reconstruct the soil stratification at the CPTu testing point.The reconstructed soil stratification has good consistency with corresponding borehole results,and the consistency level is about 95%.The RF algorithm shows optimal performance for solving this imbalance classification problem.