This paper presents a Knowledge-Constrained Machine Learning (KCML) framework to enhance yield modeling and understanding in mixed rice cropping systems, addressing the challenge of disentangling single- and double-season rice yields within these areas. The complexity of planting structures, diverse cropping calendars, and unknown genetic coefficients limits the effectiveness of process-based crop models, while conventional machine learning are hindered by the data scarcity of season-specific yield observations, as most statistical records provide only aggregated regional averages. To overcome these challenges, we integrate satellite-derived knowledge of cropping intensity into the Extreme Gradient Boosting (XGBoost) model to constrain the learning objective function and feature space. Experimental results from a typical mixed cropping region in South China demonstrate that KCML outperforms conventional ML models, achieving an overall correlation coefficient (r) of 0.72, RMSE of 934.7 kg ha⁻1(13.1%), and bias of −1.2%. Specifically, KCML successfully disentangles the yields of single and double rice, reducing the underestimation of single rice yields (bias from -9.2% to -2.6%, RMSE from 1236.8 kg ha⁻1 to 1088.8 kg ha⁻1) and overestimation of double rice yields (bias from 3.7% to 1.4%, RMSE from 713.8 kg ha⁻1 to 651.6 kg ha⁻1). The proposed KCML framework provides a unified approach for yield estimation across heterogeneous cropping systems, bridging theoretical advancements in intelligent computing with practical applications in precision agriculture.