An Integrated Approach Remote Sensing Gis Engineering Data Mining For Modeling Assessing And Mitigating Slope Stability Hazards In Mountainous Environments

Author : Te Pei
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (14 download)

DOWNLOAD NOW!

Download or read book Integrating Geotechnical Domain Knowledge Into Machine Learning for Slope Stability Predictions written by Te Pei and published by . This book was released on 2023 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: The rise of artificial intelligence (AI) and machine learning (ML) is revolutionizing various industries, including geotechnical engineering. These innovative techniques offer promising avenues for geohazard prediction, boasting the ability to learn from and make decisions based on patterns in the data. However, data-driven ML models have their inherent limitations and, especially when trained with limited data, can generate counter-intuitive or physically inconsistent results, hampering their interpretability, reliability, and overall applicability in risk forecasting for geotechnical engineering. Integrating domain knowledge into data-driven models has been widely used in many disciplines, such as computer vision, natural language processing, healthcare, criminal recidivism, and finance; however, the adoption of data-driven models with integrated domain knowledge in geotechnical engineering is still limited. This dissertation presents an innovative integration of geotechnical domain knowledge and ML techniques, addressing the unique challenges in the geotechnical engineering field, particularly in slope stability prediction and landslide susceptibility mapping. The study first examines three easy-to-implement and effective methods to integrate geotechnical domain knowledge into data-driven models for slope stability prediction: hybrid modeling, knowledge-based model initiation, and knowledge-guided loss function. A compilation of slope stability case histories from the literature was used as the benchmark database to train and validate the proposed approach. Results showed that the three proposed methods displayed superior performance to both domain knowledge-based and purely data-driven models. Notably, the hybrid-knowledge-data models and the knowledge-guided loss function managed to diminish discrepancies in slope stability predictions compared to the reported factor of safety values, resulting in predictions more consistent with the underlying physics of slope stability. Based on these findings, the second study addresses another challenge in ML application within geotechnical engineering: the tendency of ML models to produce counter-intuitive predictions when trained with limited data. A monotonicity constraint was proposed to regularize model response with respect to changes in input features, and results were compared with several benchmark ML models using case histories. The results showed that unconstrained ML models often generate predictions that violate monotonicity in many parts of the input space. However, introducing monotonicity constraints to ML models effectively mitigated these violations while maintaining high-performance levels. Consequently, this resulted in more robust and interpretable predictions as input features for geotechnical engineering applications often represent physical parameters following intrinsic and often monotonic relationships. Furthermore, the effectiveness of the proposed framework is further evaluated in a case study of landslide susceptibility mapping. A physics-guided machine learning (PGML) model that integrates outputs from a physics-based infinite slope model was applied to estimate landslide risk in the Colorado Front Range using a well-documented inventory of debris flows triggered by a single storm event. The spatial cross-validation approach was used to ensure a realistic evaluation of the model's generalization capabilities across varying ecoregions. Results showed that the proposed PGML model enhanced prediction accuracy, improved physics consistency, and reduced uncertainties compared with benchmark ML models. This underscores the limitations of pure data-driven models for predicting landslide risk across heterogeneous environments and emphasizes the essential role of incorporating geotechnical domain knowledge into ML models. In summary, this dissertation establishes a solid foundation for integrating geotechnical domain knowledge into ML models, enabling more reliable and robust predictions for geotechnical engineering applications. It underscores the importance of future research to continually explore and refine such integrated models in light of challenges associated with data limitations and regional variances.