Accurate fresh weight (FW) estimation is essential for growth monitoring and yield prediction in greenhouse fruit vegetables, but remains challenging due to the dynamic allocation between vegetative and reproductive organs. This study aimed to systematically evaluate modeling strategies for FW estimation in sweet pepper and identify which approach is most suitable under conditions of dynamic biomass partitioning. Non-destructive morphological measurements were collected under greenhouse cultivation, and allometric models based on geometric equations were established as baselines. Their performance was compared with machine learning (ML) models and ensemble learning frameworks. To address limited data availability, numerical data augmentation with Gaussian noise and a variational autoencoder was applied. Among the allometric models, the stick model combined with a sigmoid function showed the highest performance, with an $R^2$ of 0.80 for shoot FW and 0.54 for fruit FW. All ML models outperformed the allometric models, and the ensemble model achieved the highest predictive accuracy, with an $R^2$ of 0.96 for shoot FW and 0.89 for fruit FW. Data augmentation further improved predictive performance across all ML models, particularly for fruit FW prediction. Feature contribution analysis revealed that temporal progression was the dominant predictor of fruit FW, while structural traits played the primary role in shoot FW estimation. Ensemble-based ML, combined with data augmentation, provides a methodological framework for non-destructive FW estimation of sweet pepper in controlled environments such as greenhouses and smart farming systems.