Uncertainty-guided U-Net for soil boundary segmentation using Monte Carlo dropout

Accurate soil stratification is essential for geotechnical engineering design. Owing to its effectiveness and efficiency, the cone penetration test (CPT) has been widely applied for subsurface stratigraphy, which relies heavily on empiricism for correlations to soil type. Recently, deep learning techniques have shown great promise in learning the relationship between CPT data and soil boundaries automatically. However, the segmentation of soil boundaries is fraught with model and measurement uncertainty. This paper introduces an uncertainty-guided U((-Net (UGU-Net) for improved soil boundary segmentation. The UGU-Net consists of three parts: (a) a Bayesian U-Net to predict a pixel-level uncertainty map, (b) reinforcement of original labels on the basis of the predicted uncertainty map, and (c) a traditional deterministic U-Net, which is applied to the reinforced labels for final soil boundary segmentation. The results show that the proposed UGU-Net outperforms the existing methods in terms of both high accuracy and low uncertainty. A sensitivity study is also conducted to explore the influence of key model parameters on model performance. The proposed method is validated by comparing the predicted subsurface profile with benchmark profiles. The code for this project is available at github.com/Xiaoqi-Zhou-suda/UGU-Net.