Context Timely and accurate estimation of nitrogen content and aboveground biomass is essential for assessing alfalfa growth and diagnosing nitrogen deficiency. Research on efficient monitoring of growth indicators using multitask learning combined with multisource information remains limited. Aims This study aims to predict alfalfa growth indicators using machine learning and multitask learning based on multisource data, such as unmanned aerial vehicle (UAV) remote sensing data, meteorological data, and management factors, to evaluate and improve the accuracy of prediction. Methods Field-collected growth indicators including leaf nitrogen concentration (LNC), plant nitrogen concentration (PNC), aboveground biomass (AGB), and the above features were analyzed alongside feature selection based on Boruta algorithm. Machine learning was developed using random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), and multitask learning based on deep neural network (MTL-DNN), with consideration given to the contribution of features to indicators. Conclusion The results showed that prediction accuracy of alfalfa growth indicators could be significantly improved by using multisource data. RF achieved the highest accuracy for nitrogen indicators (LNC: R2 = 0.60, RMSE = 0.51%, NRMSE = 10.66%; PNC: R2=0.70, RMSE = 0.40%, NRMSE = 11.06%), The prediction accuracy for plant-level nitrogen content was consistently higher than that for leaf-level nitrogen content. Meanwhile, XGBoost performed best for AGB prediction with multisource data (R2 = 0.81, RMSE = 0.66 t ha-1, NRMSE = 16.85%). The MTL-DNN model enabled accurate and efficient prediction of alfalfa indicators, achieving R2 values of 0.66, 0.76, and 0.84 for LNC, PNC, and AGB, respectively. The AGB prediction showed the highest precision (RMSE = 0.61 t ha-1, NRMSE = 15.54%). The MTL-DNN model reduced computation time by 37.4–72.9% compared to single-task model. Implications and Impacts: This study provides a theoretical and technical basis for nutrient diagnosis, biomass estimation, and precision management of alfalfa. Implications and Impacts This study provides a theoretical and technical basis for nutrient diagnosis, biomass estimation, and precision management of alfalfa.

Multisource data and multitask deep learning for predicting alfalfa growth indicators

期刊论文
发表日期：

2026年3月27日

查看全文

Jiali Guo,
He Zhao,
Haibin Tan,
Yunling Wang,
Haijun Yan

摘要

Context

Timely and accurate estimation of nitrogen content and aboveground biomass is essential for assessing alfalfa growth and diagnosing nitrogen deficiency. Research on efficient monitoring of growth indicators using multitask learning combined with multisource information remains limited.

Aims

This study aims to predict alfalfa growth indicators using machine learning and multitask learning based on multisource data, such as unmanned aerial vehicle (UAV) remote sensing data, meteorological data, and management factors, to evaluate and improve the accuracy of prediction.

Methods

Field-collected growth indicators including leaf nitrogen concentration (LNC), plant nitrogen concentration (PNC), aboveground biomass (AGB), and the above features were analyzed alongside feature selection based on Boruta algorithm. Machine learning was developed using random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), and multitask learning based on deep neural network (MTL-DNN), with consideration given to the contribution of features to indicators.

Conclusion

The results showed that prediction accuracy of alfalfa growth indicators could be significantly improved by using multisource data. RF achieved the highest accuracy for nitrogen indicators (LNC: R² = 0.60, RMSE = 0.51%, NRMSE = 10.66%; PNC: R²=0.70, RMSE = 0.40%, NRMSE = 11.06%), The prediction accuracy for plant-level nitrogen content was consistently higher than that for leaf-level nitrogen content. Meanwhile, XGBoost performed best for AGB prediction with multisource data (R² = 0.81, RMSE = 0.66 t ha^-1, NRMSE = 16.85%). The MTL-DNN model enabled accurate and efficient prediction of alfalfa indicators, achieving R² values of 0.66, 0.76, and 0.84 for LNC, PNC, and AGB, respectively. The AGB prediction showed the highest precision (RMSE = 0.61 t ha^-1, NRMSE = 15.54%). The MTL-DNN model reduced computation time by 37.4–72.9% compared to single-task model. Implications and Impacts: This study provides a theoretical and technical basis for nutrient diagnosis, biomass estimation, and precision management of alfalfa.

Implications and Impacts

This study provides a theoretical and technical basis for nutrient diagnosis, biomass estimation, and precision management of alfalfa.