Multi-Source Environmental Information Fusion and Adaptive Deep Learning for Karst Landslide Displacement Prediction

To address the challenges of information fusion and prediction for highly non-stationary, noisy, and lag-responsive heterogeneous time-series data from multi-source environmental sensing, this paper proposes a novel adaptive hybrid framework, SAPSO-VMD-GRU. First, the framework employs Variational Mode Decomposition (VMD) to decouple the target sequence into trend, periodic, and random components to reduce complexity and filter noise. Then, Lagged Cross-Correlation Analysis (LCCA) is introduced to quantify the time-lagged correlation between the target components and variables such as rainfall and multi-depth soil temperature and moisture, eliminating redundant features to achieve deep fusion of multi-source information. This paper designs an adaptive particle swarm optimization algorithm, SAPSO, by integrating improved Circle chaotic initialization, Sa-function-based nonlinear inertia weight, and a two-stage Cauchy mutation strategy. SAPSO is used to adaptively determine the VMD parameters and the key GRU hyperparameters in different modeling stages. Experiments based on the Bayintun landslide dataset show that, by utilizing the past 5 days of multi-source historical data, including GNSS displacement, rainfall, and lagged soil moisture and temperature, as inputs to forecast the next-day displacement, the proposed framework achieved R2 values above 0.95 on the chronological hold-out validation subset at all three GNSS monitoring stations. These results indicate that the proposed framework can effectively capture lagged triggering effects and improve displacement prediction accuracy under complex, noisy, and non-stationary monitoring conditions.