Physics-Informed Multimodal Graph-Temporal Network with Agentic Hydro-Mechanical Verification for Dynamic Slope Stability Prediction under Rainfall-Vegetation-Soil Coupling

Overview Rainfall-induced shallow landslides on vegetated slopes are governed by nonlinear feedbacks among canopy interception, root water uptake, capillary hysteresis, and moisture-dependent root tensile reinforcement. Purely data-driven sequence models frequently violate water mass conservation and matric suction monotonicity during convective cloudbursts, whereas static soft-penalty physics-informed neural networks overlook drying-wetting retention loops, near-saturation root-soil slippage, and preferential macropore bypass. We present PIM-RVSNet, a physics-informed multimodal graph-temporal network coupled with an agentic hydro-mechanical equilibrium verifier for dynamic Factor of Safety (F_s(t)) and instability lead-time prediction. PIM-RVSNet integrates three irreducible mechanisms: a Multimodal Cross-Scale Eco-Hydrological Encoder (MC-EHE) using multi-head Depth-Directed Eco-Hydrological Cross-Attention (DEH-CA) and vertical soil-horizon graph-recurrent message passing to align atmospheric forcing, depth-specific canopy-root traits, and four-layer soil profiles; a Differentiable Richards-Wu-Mohr-Coulomb Coupled Operator (RWMC-Op) embedding scanning-curve van Genuchten hysteresis and saturation-attenuated root cohesion; and an Agentic Hydro-Mechanical Equilibrium Verifier (AHM-Verifier) that audits mass-balance drift, preferential bypass, and stress admissibility to iteratively recalibrate upper-boundary infiltration partitioning within at most three reflexion steps. Across 48 monitored slope units and 240 storm episodes (2,880 space-time states across four eco-geomorphic zones), PIM-RVSNet achieves a test coefficient of determination R^2 of 0.9514 ± 0.0028, RMSE_F_s of 0.1106 ± 0.0032, critical instability F1_crit of 0.8967 ± 0.0168, and a physical violation rate of 1.42 ± 0.41%, outperforming the strongest physics-informed baseline (R^2 = 0.9158 ± 0.0032, p_Holm = 1.24 × 10^-13) and extending mean pre-onset early-warning lead time from 2.15 h to 3.98 h. Contents This record contains 15 files. The table lists every data table in the archive with its row count. FileRowsSize (KB)data/benchmark/episode_timeseries.csv2,880599data/benchmark/slope_units.csv486data/derived/ablation_table.csv102data/derived/component_replacement.csv101data/derived/dataset_summary.csv40data/derived/error_diagnostics.csv242data/derived/horizon_efficiency.csv302data/derived/main_table.csv163data/derived/main_table_seeds.csv809data/derived/regime_robustness.csv606data/derived/sensitivity_table.csv192data/predictions/horizon_predictions.csv108,0006858data/predictions/seed_predictions.csv108,0009449data/predictions/sensitivity_predictions.csv68,4006173 Reuse Released under cc-by-4.0. If you use these data, please cite this record together with the associated article.