Rainfall Memory Flood Index (RMFI): An I30-Based Framework for Spatiotemporal Flood Susceptibility Assessment

Flood susceptibility assessments commonly rely on static terrain variables and temporally aggregated rainfall indicators, although flood generation is strongly influenced by antecedent hydrological conditions and the seasonal structure of rainfall forcing. This study introduces the Rainfall Memory Flood Index (RMFI), an I30-based framework for spatiotemporal flood susceptibility assessment. The approach integrates elevation with a rainfall-memory component derived from monthly erosivity statistics. For each month, the rainfall-memory term is calculated from monthly mean and maximum event erosivity values obtained from rainfall events separated according to the RUSLE-R procedure, while monthly I30 is defined from the event producing the monthly maximum erosivity. Flood occurrence is modelled as a binary outcome using a 1209-point dataset comprising 65.9% flood and 34.1% non-flood observations. Two predictor structures are evaluated: an annual formulation based on elevation and total_I30 × S_total, and monthly formulations based on elevation and I30_m × S_m. Four machine-learning algorithms, Logistic Regression, Random Forest, XGBoost, and LightGBM, are evaluated using 70/30 holdout validation and stratified 5-fold cross-validation. Results show that monthly RMFI formulations consistently outperform the annual aggregation. The strongest performance is obtained in June, with cross-validated values of ROC AUC ≈ 0.86, F1 ≈ 0.84, and PR AUC ≈ 0.93, while similarly strong results are observed in January and November. These findings indicate that flood susceptibility is more effectively captured when rainfall forcing is represented through a temporally explicit, memory-sensitive structure rather than through annual accumulation alone, supporting a parsimonious, physically interpretable, and transferable framework for flood susceptibility mapping in data-constrained environments.