The integration of digitization and automation into traditional farming has transformed agriculture, enhanced productivity and delivered significant economic benefits. Recent advances in deep learning (DL) provide new opportunities for automating crop monitoring and nutrient management. This paper proposes a framework for estimating nitrogen content in maize leaves using RGB images captured under field conditions.
A UNet-based segmentation model isolates leaves from complex, uncontrolled backgrounds, and a convolutional neural network (CNN) extracts discriminative features. These features are evaluated using three regression models, Support Vector Regressor (SVR), Random Forest Regressor (RFR), and Gradient Boosting Regressor (GBR), to predict SPAD values, ranges from 24.9 to 59.5, an indirect measure of chlorophyll and nitrogen content.
Experiments conducted on six maize varieties (three open-pollinated and three hybrid) across four growth stages yielded an overall lowest mean absolute error (MAE) of 3.56 and a correlation coefficient of 0.76. Aggregating results by dataset type, the best-performing method achieved an average MAE of 3.56 for OPV and 4.04 for Hybrid varieties.
Further analysis by individual variety type revealed that RFR was most effective, attaining the lowest MAEs of 1.34 and 1.47 for OPV and Hybrid varieties, respectively. A second contribution is MaizeRGB, a comprehensive, publicly available dataset containing RGB images and ground-truth SPAD measurements for multiple maize varieties, nitrogen levels, and growth stages.
The proposed approach demonstrates high accuracy, robustness to variable field conditions, and strong potential for low-cost, scalable nitrogen management in precision agriculture.