Rapid and accurate monitoring of soil total nitrogen (STN) is fundamental to optimising nitrogen fertiliser management, improving nitrogen use efficiency, and reducing agricultural non-point source pollution; however, traditional methods struggle to provide timely, large-scale, and continuous monitoring. Based on the relationship between maize canopy spectral characteristics and soil nitrogen distribution, this study acquired multi-spectral drone imagery during the jointing, tasselling, and grain filling stages, whilst simultaneously collecting soil samples from the 0–20, 20–40, and 40–60 cm soil layers. Using canopy reflectance (Band), vegetation indices (VIs), textural features (TF), and a fusion of these three features (Band+VIs+TF) as input variables, machine learning models, including Random Forest Regression (RFR), Gradient Boosted Decision Trees (GBDT), K-Nearest Neighbours (KNN), and XGBoost, were constructed to estimate STN. The results indicate that (1) STN is significantly correlated with spectral reflectance, vegetation indices, and textural features, with absolute correlation coefficients ranging from 0.60 to 0.89; (2) the accuracy of STN estimation by the four models showed a decreasing trend with increasing soil depth, with the highest accuracy observed in the 0–20 cm soil layer; (3) the validation accuracy of STN estimation followed the order: grain filling stage (R² = 0.833) > tasselling stage (R² = 0.794) > jointing stage (R² = 0.761); (4) under feature fusion conditions, the STN estimation accuracy of all four models was superior to that of using a single feature, and the prediction accuracy of the XGBoost model (0.833) was higher than that of the other three models (0.701–0.734). In summary, during the grain filling stage of maize, using feature fusion as input variables and the XGBoost model can significantly improve the accuracy of STN estimation in maize fields, with the 0–20 cm soil layer yielding the best results.