Abstract: Regional landslide early warning systems play a crucial role in disaster prevention and mitigation in China. In recent years,the risk of clustered landslides triggered by extreme rainfall has increased,highlighting the need for refined early warning to support timely evacuation and risk reduction. Previous machine learning-based early warning models have commonly used grid cells or slope units as evaluation units for computational convenience. However,systematic comparisons of the performance and practical effectiveness of these two evaluation units in machine learning models remain limited. This study takes Qingchuan County,Sichuan Province,as the research area and employs the same algorithm(Random Forest)and training dataset(nine years of records)to construct regional landslide early warning models based on grid cells and slope units,respectively. A systematic comparison was conducted from three perspectives: quantification of landslide influencing factors,model performance evaluation,and case validation. The analysis indicates that slope units outperform grid cells in reflecting key landslide influencing factors,model discrimination capability,and the spatiotemporal resolution of warning results,enabling more accurate prediction of regional landslide distribution. Regional landslide early warning models based on slope units demonstrate clear advantages for refined warning applications,providing valuable insights for landslide risk prevention and mitigation,and supporting the adoption of slope units in future studies on refined landslide early warning.