Predicting Soil Behavior Types Along the Danube: An AI-Driven Approach Using CPT Data in the Szigetköz Floodplain Area

Abstract

The Szigetköz (Hungary) is a hotbed of sand boil formation, owing to the combination of a 100-250 m thick gravel layer beneath a relatively thin covering of poor soil with varying thickness. Soil behavior is critical for flood protection in this region. This work proposes a novel way to predict Soil Behaviour Types (SBT) based on detailed CPT data collected from 29 sites in the Szigetköz area using an artificial intelligence (AI) model. The study follows a methodically planned approach that includes data collecting, preprocessing, SBT categorization based on the SBT chart developed by Robertson et al. (1986), and AI model building. The CPT dataset contains critical metrics like cone resistance and friction ratio, which are essential in characterising soil behavior. The AI model, built with powerful machine learning algorithms, is intended to learn complicated associations within data to forecast SBT classifications. Extensive feature selection, hyperparameter tuning, and cross-validation are all necessary steps in model construction to ensure accuracy and generalizability. The results show that the model can accurately forecast SBT classifications for the Szigetköz area, shedding information on the soil's behavior near the Danube River. Spatial distribution visualizations emphasize the region's many SBT categories, giving valuable information for engineering projects, land use planning, and environmental conservation activities. The AI model's interpretability elucidates the major CPT parameters driving SBT forecasts, providing stakeholders with actionable information for decision-making. Furthermore, validation of the model with new, previously unseen CPT data confirms its applicability and robustness in real-world circumstances.

Full Paper