Displacement prediction for long-span bridges via limited remote sensing images: An adaptive ensemble regression method

• Proposing a novel predictive method based on an adaptive ensemble regression algorithm. • Leveraging advanced ML algorithms such as kernel learning, ensemble learning, incremental learning, and hybrid learning. • Simultaneous predicting and normalizing using limited data. • Validating real-world data of long-span bridge. Spaceborne remote sensing via synthetic aperture radar (SAR) images offers promising solutions to long-term structural health monitoring by providing local displacement time histories. However, this methodology faces challenges such as limited image accessibility, data sparsity, and real-time monitoring feasibility. Although regression-based prediction is a practical approach to deal with these limitations, the availability of limited SAR-extracted displacement data and the impacts of unmeasured environmental/operational factors lead to extra challenges that can skew prediction outputs. To overcome these issues, this article proposes a novel adaptive ensemble regression method that not only predicts displacement time series from limited SAR images but also simultaneously removes environmental/operational variability in predicted displacements. This method features two levels of kernelized and adaptive regression modeling within a sequential ensemble learning framework using Gaussian process regression as the primary regressor. Results from two real-world bridge structures substantiate the effectiveness of the proposed method in simultaneous prediction and normalization.