Independent Multi-Sensor Validation of Machine-Learning Landslide Susceptibility: Footprint Construction Decides the Verdict—May 2023 Emilia-Romagna Event

Machine-learning landslide-susceptibility maps are almost always judged by inventory-split skill (the area under the receiver-operating-characteristic curve, AUC, and Cohen’s κ), not by model-independent physical observation of where an event caused ground disturbance. For the May 2023 Emilia-Romagna event (>80,000 landslides; RER2023 inventory), we confront an open-data, event-conditioned susceptibility model (trigger rainfall is among its predictors) with a co-event disturbance footprint built from two satellites: phenology-matched Sentinel-2 change in the Normalized Difference Vegetation Index (ΔNDVI) and Normalized Burn Ratio (ΔNBR), and a 12-day Sentinel-1A C-band coherence-and-backscatter layer used as a cloud-independent coverage check (C-band 12-day decorrelation is the a priori expectation in this setting); neither enters the model. Apparent geomorphic plausibility depends critically on how the independent footprint is constructed. Against a raw co-event footprint (contaminated by flooding and agriculture), a model with AUC ≈ 0.945 is indistinguishable from a random mask. On a landslide-relevant footprint, the same model captures optically detectable disturbance at roughly twice the chance (bootstrap median 1.94×, 95% CI 1.25–2.72, excluding 1; ≈1.46× above a vegetation-phenology null), but the aggregate is driven by an upper-tail minority of 10 km blocks—the majority of blocks have per-block median lift 0.67× (below chance). The map is therefore a regional, not a per-place, statement. Results are a within-event consistency test; cross-event transferability is not claimed. Footprint construction is decisive and currently neglected. The open-source pipeline is released upon acceptance.