Deep Learning Characterization of Spatially Variable Ground Using Screened and Label-Conditioned Random Field Data

<title>Abstract</title> <p>The scale of fluctuation (SoF) is a key statistical parameter for characterizing the spatial variability of geotechnical properties. Recent studies have employed Convolutional Neural Networks (CNNs) to estimate the SoF using synthetic random field realizations generated through LU decomposition-based random field theory. However, these datasets may contain non-stationary realizations and inconsistencies between the prescribed SoF values and the generated random fields, resulting in unreliable training labels. The present study proposes a systematic framework for constructing reliable CNN training datasets by integrating non-stationarity screening, Maximum Likelihood Estimation (MLE)-based verification of SoF labels, and a label-conditioning strategy to establish physically consistent realization-SoF mappings. Four datasets with different levels of screening and conditioning are developed to evaluate the influence of dataset quality on CNN performance. The proposed framework is validated using independent Cone Penetration Test (CPT) profiles and compared with a baseline CNN trained on the unprocessed dataset. The results demonstrate that the proposed framework significantly improves prediction accuracy, learning stability, and generalization to field data. The study highlights that the reliability of CNN-based characterization of spatially variable ground depends not only on the network architecture but also on the quality and physical consistency of the training dataset.</p>