Characterizing Seasonality and Trend From In Situ Time‐Series Observations Using Explainable Deep Learning for Ground Deformation Forecasting

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 2
• Main Authors: Ma, Zhengjing, Mei, Gang, Xu, Nengxiong
• Format: Journal Article
• Language: English
• Published: 01.06.2024
• Subjects: deep learning explainable neural network; geohazards; ground deformation forecasting; seasonality; time‐series observations; trend
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000122

Ground deformation, a critical indicator of geohazard evolution, exhibits both seasonal fluctuations and long‐term trend changes. This study explores interpretable deformation forecasting using an explainable deep learning approach, utilizing field observation data to extract and characterize these critical features from time series. We demonstrate that extracting and utilizing shared, similar seasonal and trend features across multi‐point observations significantly enhances ground deformation forecasting accuracy. Notably, the extracted seasonal features reveal the dominant role of the hydrological response in deformation at seasonal timescales, providing insights into the underlying source of predictability. This proof‐of‐concept study highlights the promise of interpretable ground deformation forecasting, with applications expected to encompass landslide behavior, volcanism, or seismic activity. Plain Language Summary Monitoring ground deformation is critical for forecasting geological hazards such as landslides. However, ground deformation forecasting can be challenging due to its unpredictable and complex nature. An interpretable ground deformation forecasting paradigm has been presented here, in which (a) explainable neural networks are used with prior constraints related to seasonal and trend features, and (b) multitask learning is integrated across multiple observation sites to capture related seasonal and trend patterns. We evaluated this approach using data from landslide monitoring systems spanning diverse deformation patterns and observation intervals. The results demonstrate the method's ability to enhance forecasting performance by characterizing seasonal and trend patterns across multiple observation sites. As an initial demonstration, this study explores the potential of interpretable forecasting approaches, which could be applied to forecast ground deformation patterns on various geological hazards. In the long run, the interpretable forecasting paradigm is promising for understanding and forecasting natural hazards. Key Points An explainable neural network reveals seasonal and trend patterns in ground deformation observations Understanding these patterns reveals the underlying deformation processes caused by external factors Utilizing similar seasonal and trend patterns across observational sites improves ground deformation forecasting