Abstract

Nitrogen deficiency remains one of the most critical constraints to sustain maize productivity. While conventional measurement techniques are destructive, costly, and labor-intensive; therefore, a non-destructive, smartphone-based framework that integrates automated leaf segmentation using a deep learning model (U²-Net) with color-metric analysis to estimate nitrogen content directly from field-acquired images is developed. The model was trained and evaluated on a comprehensive dataset of 1,188 maize (Zea mays L.) leaf samples from 11 varieties, cultivated under varying nitrogen treatments and photographed using multiple smartphones. Estimation results showed a significant correlation with both laboratory Kjeldahl digestion benchmarks and Greenseeker normalized difference vegetation index (NDVI) readings. Collective and variety-specific evaluations demonstrated strong agreement with reference measurements, with correlations reaching ρ = 0.65 overall and up to ρ = 0.95 at the variety level. Among the 26 tested color indices, B/(R+G+B), R-G and G-B consistently emerged as the most reliable predictors of leaf nitrogen. Statistical testing using PERMANOVA further revealed that camera resolution significantly influences several color metrics, emphasizing the importance of device-aware calibration in practical deployments. This study demonstrates the potential of smartphones as scalable tools for nitrogen monitoring by combining accessibility, accuracy, and reproducibility, and also contributes an open dataset to support future research in digital agriculture. Future model iterations will aim to improve robustness to challenging field variables like extreme illumination, shadows, other camera resolutions and diverse leaf orientations, along with investigation of applicability with other cereal crops.