Abstract: Aiming at the precision bottleneck and insufficient spatial heterogeneity analysis in existing models due to single-algorithm dependence,this study proposes a three-tier “multimodal ensemble-parameter adaptation-feature enhancement” optimization framework.First,multi-source data (such as nighttime lighting,building outlines)are integrated to construct a secondary model (N-MLP)via stacking random forest,XGBoost,and MLP.Then,the IVY algorithm (IVYA)is introduced for dynamic hyperparameter optimization,and a spatial interaction-augmented attention mechanism (SIAM)is designed to enhance geographical spatial dependence analysis through parallel attention architectures.Finally,a dual-scale validation system (400 m grid and township/street levels)is established in Chengdu,and the driving effect of population distribution on drone logistics demand is analyzed via a low-altitude economy demand elasticity model.Results show that the optimized SIAM-IVYA-N-MLP model achieves an R² of 0.947 9 at the grid scale,with MAE and RMSE reduced by 14.67%and 3.38%,respectively.At the township/street scale,the R² reaches 0.971 6.A 1%increase in main urban population density drives a 1.19%growth in drone logistics demand.This study provides an operational technical pathway for high-precision population spatialization and low-altitude economic infrastructure planning.

Key words: population spatialization, IVY algorithm, spatial interaction-attention mechanism, ensemble learning, low-altitude economy