The prediction of loess cave density is challenged by a fundamental scale issue. Local-scale studies yield high-resolution distribution maps but generalize poorly. Conversely, regional assessments identify broad patterns but tend to overlook the local topographic and soil heterogeneities that control loess cave development. The gully catchment scale resolves this problem by integrating hillslope-channel processes at a scale suitable for both mechanistic analysis and regional application. Analysis of 44 gully catchments across the western Loess Plateau using UAV surveys, soil sampling, and climate data quantified how hydrogeomorphic, soil, and climatic factors control spatial loess cave density variations. A Normalized Process-informed model was subsequently developed, incorporating these controls through a coupled non-linear framework. Leave-one-site-out cross-validation confirmed that the proposed model exhibits superior transferability, consistently outperforming traditional linear and power-law regressions, as well as machine learning benchmarks including Random Forest and XGBoost. In the study area, loess cave abundance peaks in silt-dominated areas where gully catchments are characterized by a poor mean index of connectivity and limited TVDI. When applied regionally, the model captured the northwest-to-southeast increasing density gradient documented in field surveys, demonstrating its ability beyond the original calibration area. By operating at the gully catchment scale, this approach offers a practical tool for loess cave assessment, combining process understanding with predictive capability, while retaining the physical basis needed for erosion hazard evaluation and land management decisions.