Abstract: With the artificial propagation of Chinese sturgeon (Acipenser sinensis) surpassing one million juveniles, industrial rearing increasingly demands automated feeding and precision management. To reduce feed waste, alleviate water-treatment load, reduce manual labour intensity and costs, and enable precise on-demand feeding, this study developed a predictive model for the daily satiation feeding rate of juvenile Chinese sturgeon. Rearing data were collected over 98 days from juvenile Acipenser sinensis under recirculating aquaculture system (RAS) conditions. Using six core factors, including daily mean water temperature and age, as inputs and daily satiation feeding rate as the output, four baseline models were constructed: multiple linear regression (MLR), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). Generalization performance was evaluated using tank-based 3-fold cross-spatial cross-validation, with hyperparameters uniformly tuned via Bayesian optimization. The results showed that the tree-ensemble models (RF and XGBoost) outperformed both the linear model (MLR) and the kernel method (SVM), each achieving a test-set coefficient of determination (R²) of 0.47. Given that XGBoost offers a richer set of tunable hyperparameters and consequent greater potential for targeted improvement, XGBoost was adopted as the base model. Targeting the small-sample, high-noise, and cross-spatial distributional characteristics of the sturgeon rearing data, four improvements were integrated, namely out-of-fold (OOF) residual anomaly cleaning, metabolic-index feature construction, K-means micro-ecological clustering, and Yeo-Johnson target transformation, to propose the SCB-XGBoost (Spatial-Clustering-Bayesian XGBoost) model, in which the metabolic-index and micro-ecological clustering features expanded the model input from the original six dimensions to eight. SCB-XGBoost achieved a test-set R² of 0.53, root mean square error (RMSE) of 0.84, and mean absolute error (MAE) of 0.67, representing an improvement of approximately 13% in R² over the baseline XGBoost. Comparative experiments showed that the four improvements were effective across multiple base models but, under small-sample conditions, required a tree-ensemble model combining nonlinear fitting capability with resistance to overfitting to be fully realized; between the comparably performing SCB-RF and SCB-XGBoost, the latter exhibited a smaller model size (0.19 MB) and faster inference speed (1.2ms), rendering it more suitable for edge deployment. SHAP attribution analysis indicated that, among the eight input features, age and stocking density were the two most influential factors: the effect of age first increased and then decreased over the course of development, while stocking density exhibited a non-monotonic pattern, promoting feeding within a suitable range but suppressing it at extremes. The interaction between water temperature and age (the metabolic index) contributed more than water temperature alone. These quantitative results are consistent with the metabolic rhythm and long-term developmental dynamics of the Chinese sturgeon as a sub-cold-water species. This study provides technical support for precise feeding and digital management of juvenile Chinese sturgeon under recirculating aquaculture conditions.