Data Test of a Hydro-Kinematic State Graph for Landslide Forecasting

MOHIT SHEODE 

A graph model can make landslide-monitoring relationships explicit, but a graph is not automatically more informative than a well-regularized time-series model. This study tests that distinction on the publicly released Lamosano landslide record from northeastern Italy. The processed dataset contains 161 observations of 11-day horizontal InSAR differential displacement and rainfall from 10 April 2015 to 3 February 2020. Each graph node represents a causal hydro-kinematic state assembled from recent displacement, rainfall, trend, variability, and antecedent rainfall. Edges link a query state only to similar states in the preceding training record; the next displacement is estimated from their distance-weighted successors. Lookback and neighborhood size were selected by three blocked validation folds within the first 100 observations. The final 61 observations were held out chronologically. Climatology, persistence, and Ridge-ARX were tested under the same boundary. The selected graph produced a test RMSE of 2.451 mm and standard R² of −0.001, compared with 2.400 mm and 0.039 for Ridge-ARX. A 5,000-replicate moving-block bootstrap placed the graph-minus-ridge RMSE difference between −0.084 and 0.200 mm. Neither model recalled any of seven high-motion intervals. Thus, the graph exposed recurring monitoring states but did not supply dependable incremental forecast skill. The contribution is an auditable negative benchmark: with one monitoring point and a short record, state-graph complexity should not be interpreted as spatial knowledge or operational warning capability.