Abstract： Flash floods pose severe threats due to their abruptness and spatial heterogeneity. Traditional Flash Flood Susceptibility Mapping (FFSM) relies on static factors and empirical models, which lack spatial generalization and physical constraints under frequent extreme precipitation events and rapidly changing environments. To address the urgent need for fine-grained flash flood risk management, this study proposes a three-stage progressive research framework—comprising “spatial cognition, spatial perception, and intelligent mapping”—from the perspective of risk perception. It provides an in-depth analysis of the theoretical trajectory of FFSM’s evolution from empirical modeling to hybrid intelligence. A comprehensive analysis yields the following insights: First, flash flood susceptibility is fundamentally a potential spatial response state mediating between the environmental background and the disaster event; the core of its mapping lies in evaluating the relative spatial propensity of a watershed under similar triggering conditions. Second, by constructing a binary spatial cognition structure consisting of an “intrinsic environment” and “extrinsic triggers”, this study clarifies the observable geospatial elements in flash flood risk mapping. This structure broadens the application dimensions of space-air-ground integrated multi-source remote sensing observations in the representation of FFSM factors. Furthermore, the evolution of susceptibility mapping methods reflects an elevation in the dimensions of spatial perception. Geospatial Artificial Intelligence (GeoAI) is profoundly driving the transition of mapping research from surface-level “factor-driven” paradigms to the exploration of deep “nonlinear spatial associations”. Finally, this study systematically explores a hybrid intelligent mapping framework that integrates the implicit representational capabilities of GeoAI with physical process constraints. It highlights that establishing a synergistic mechanism between data-driven approaches and physical conservation laws is a crucial pathway for enhancing the robustness and interpretability of mapping under complex scenarios. Ultimately, this study aims to propel the paradigm shift in flash flood susceptibility mapping from experience-oriented to cognition-driven and precise intelligent governance, thereby providing robust scientific support for fine-grained disaster prevention and mitigation in mountainous cities.