Above-ground biomass (AGB) is an important indicator for assessing crop growth and guiding agricultural management. Accurately estimating potato AGB is crucial for growth monitoring and yield prediction. Traditional methods of measuring AGB cannot meet the need for large-scale, rapid monitoring. Remote sensing technology is quick, efficient, and non-destructive, making it a valuable tool for crop monitoring. However, remote sensing vegetation indices lose sensitivity in areas with high vegetation coverage, leading to a “saturation effect” that limits AGB accuracy. Most previous studies focused on improving remote sensing indices for AGB estimation. However, these methods still lack clear explanations for how potato growth affects AGB changes. In this study, we analyze the growth process of potatoes and consider the characteristics of different growth stages. We use UAV multispectral data to examine how plant parts, like stems and leaves, contribute to biomass accumulation. We develop a model to track the ratio of leaf biomass (Leaf Ground Biomass, LGB) to AGB throughout the growing season, using effective accumulated temperature (Growing Degree Days, GDD) as an independent variable. By estimating LGB through remote sensing, we can indirectly estimate potato AGB. We built a dynamic AGB allocation model using field measurements. We applied three methods—correlation analysis, variable importance projection (VIP), and random forest feature importance evaluation (FIM)—to select vegetation indices related to LGB. We then used Random Forest (RF), Gaussian Process Regression (GPR), and Partial Least Squares Regression (PLSR) to estimate AGB over the entire growing season. The results showed that: (1) the best vegetation index combination, including MTVI1, MTVI2, NDVI, GNDVI, RVI, and TVI, was selected based on R, VIP, and FIM evaluations; (2) combining the AGB dynamic allocation model with machine learning improved AGB estimation accuracy; (3) the model combining AGB dynamic allocation with random forest achieved the highest accuracy, with a training set R² of 0.80, RMSE of 256.73 kg·ha^-1, and NRMSE of 9.91%. For the validation set, R² was 0.76, RMSE was 211.91 kg·ha^-1, and NRMSE was 11.46%.