Accurate monitoring and prediction of soil moisture (SM) are essential for identifying water-sensitive growth stages, optimizing irrigation strategies, and improving agricultural productivity in Hemerocallis cultivation. This study develops a machine learning framework integrating TSM640 sensor, multi-source remote sensing (Sentinel-1/2), and meteorological datasets to analyze SM dynamics across soil layers and their impact on yield. The Backpropagation Neural Network (BPNN) model demonstrated robust performance in SM estimation (R² = 0.64), with evaluation metrics of mean absolute error (MAE) = 1.89, mean bias error (MBE) = 0.91, and root mean square error (RMSE) = 2.60. The model's multi-layer architecture proved particularly effective for handling high-dimensional feature spaces. A significant 0.15 improvement in R² was achieved through the synergistic use of Sentinel-1 and Sentinel-2 data, with corresponding reductions in error metrics (MAE: −0.57, MBE: −0.81, RMSE: −0.71) compared to single-source approaches. Incorporating meteorological data showed limited predictive enhancement due to the complex timing between moisture response and meteorological conditions. For yield prediction, the Random Forest (RF) algorithm outperformed BPNN (R² = 0.63, MAE = 298.99, MBE = 78.16, RMSE = 366.84). Feature importance analysis revealed critical growth stages: Bolting (0.42), Squaring (0.32), Filling (0.28), and Seedling (0.25). Specifically, these results identify bolting and squaring phases as particularly moisture-sensitive periods, suggesting that targeted irrigation during these developmental windows could significantly improve Hemerocallis productivity.