Multi-sensor based landslide monitoring via transfer learning

Abstract Landslides are severe geographical activities that result in large quantities of rock and debris flowing down hill-slopes, leading to thousands of casualties and billions of dollars in infrastructure damage every year worldwide. For detecting landslides, on-site sensor systems are widely applied for data collection and many existing statistical process control methods can be adopted for modeling and monitoring. However, the conventional methods may perform poorly or even inapplicable when the sensors have different set-up times and end times, especially when the system includes newly deployed sensors with limited data collected. To make effective use of such new sensors immediately after deployment, we propose a novel multi-sensor based charting scheme for dynamic landslide modeling and monitoring by using transfer learning. A regularized parameter-based transfer learning approach integrated with the ordered LASSO is first proposed to effectively transfer information from old sensors with sufficient historical data to new ones with limited data. The approach considers the similarities not only between the autoregressive coefficients of different sensors, but also between the temporal correlation patterns. A control chart is then proposed for monitoring the newly deployed sensors sequentially based on the generalized likelihood ratio. Extensive simulation results and a real data example of landslide monitoring demonstrate the effectiveness of our proposed method.