基於層級貝葉斯模型之CPT土壤分類研究

Abstract

現今常見的土壤分類系統為統一土壤分類系統 (Unified Soil Classification System, USCS)。然而在現地實務中，受限於經費與工期，往往僅能於特定深度土壤取樣分類，難以獲得連續的地層資訊。近年來，圓錐貫入試驗 (cone penetration test, CPT) 因具備高度重現性與近乎連續的量測能力，逐漸成為現地調查的主流手段。CPT的一項重要應用為土壤分類，其中又以Robertson (2009) 所提出的正規土壤行為類別 (normalized soil behavior type, SBTn) 最具代表性。該方法可實現沿深度方向的連續土壤分類，有效降低USCS因樣本稀少所帶來的不確定性。 在大地工程中，另一項不可忽視的挑戰為「場址獨特性」。SBTn方法是依據大量歷史資料迴歸所建構的通用模型，雖於多數場址中表現良好，然而在特定場址上可能因場址特性差異而降低其適用性。此外，場址特定資料通常相當有限，使得建立之場址特定模型常伴隨較高的統計不確定性。 為解決上述問題，本研究首先建立一套名為「CPT-USCS/8/4182」之資料庫，整合來自全球 506 個場址共 4182 筆CPT試驗與USCS土壤分類對應資料，作為後續模型訓練之基礎。此外，本研究提出一套改良之層級貝葉斯模型 (hierarchical Bayesian model, HBM) 架構，稱為USCS-HBM，用以處理CPT土壤分類問題，並學習資料庫中各場址的場址特徵。 USCS-HBM經由資料庫訓練後，能為任意目標場址產生對應的先驗模型，該模型可進一步融合目標場址的少量觀測資料進行更新，進而推估出準場址特定模型。透過本研究所提出之 USCS-HBM 架構，能有效處理CPT土壤分類中「場址獨特性」與「資料稀疏性」所帶來的挑戰，提升分類準確性與模型實務應用性。

The Unified Soil Classification System (USCS) is one of the most commonly used soil classification methods. However, in practice, due to budget and time limits, soil sampling is often done only at specific depths. In recent years, the cone penetration test (CPT) has become a popular site investigation method because of its high repeatability and nearly continuous measurements. One key application of CPT is soil classification, with the normalized soil behavior type (SBTn) (Robertson 2009) being widely used. This method allows continuous soil classification with depth and helps reduce the uncertainty of USCS caused by limited sampling. A key challenge that should not be overlooked in geotechnical engineering is site-uniqueness. The SBTn method is a generic model constructed through regression analysis of extensive historical data. While it performs well in most global site, its applicability may be reduced at certain sites due to differences in site characteristics. Moreover, site-specific data are often limited, which introduces a higher degree of statistical uncertainty when developing site-specific models. To address these issues, this study develops a global database, CPT-USCS/8/4182, containing 4,182 CPT-USCS data pairs from 506 sites. In addition, this study proposes an improved hierarchical Bayesian model (HBM) framework, referred to as USCS-HBM, to address CPT-based soil classification and to learn site-specific characteristics from the database. The trained model can generate a prior for any target site and update it with sparse target-site data to produce a quasi-site-specific model. This approach improves CPT-based soil classification under conditions of site uniqueness and the practical challenge of sparse target-site data.