The escalating spread of the competitive weed Chenopodium album poses a substantial threat to strategic oilseed production, necessitating advanced, spatially explicit risk-assessment tools. This study develops and compares two hybrid machine-learning architectures, the RF-CatBoost stacked ensemble and the CNN-XGBoost deep ensemble, to model habitat suitability of C. album across the heterogeneous rapeseed-growing landscapes of Fars Province, Iran. Using 18 optimized environmental covariates combined with extensive field observations, both models delivered strong predictive performance (AUC >= 0.82), outperforming conventional classifiers. The RF-CatBoost hybrid demonstrated the highest accuracy (AUC = 0.84 f 0.03; TSS = 0.68 f 0.04; Kappa = 0.63 f 0.05) and the most stable spatial behavior (sigma = 0.069; CV = 8.1%), whereas the CNN-XGBoost model showed moderately lower accuracy (AUC = 0.82 f 0.04; TSS = 0.65 f 0.05) with higher spatial uncertainty (sigma = 0.078). Variable-importance analyses indicated that habitat suitability is primarily governed by edaphic factors, especially Clay content, Organic Matter, and Phosphorus, alongside anthropogenic dispersal pathways including proximity to roads and rivers. The resulting Habitat Suitability Maps (HSMs) reveal substantial areas of High and Very High suitability, accounting for 22.9% of the province under the RF-CatBoost model, thus offering a robust evidence base for Precision Weed Management (PWM). Overall, the findings demonstrate the effectiveness of deep ensemble learning in producing reliable ecological risk assessments and highlight priority zones where targeted management can enhance the sustainability and economic efficiency of rapeseed production in semiarid regions.