Image-based classification of fertilizer-induced nitrogen variation in maize leaves (Zea mays L.) through deep neural networks

Nitrogen mismanagement limits maize productivity in smallholder systems, yet rapid, low-cost nitrogen diagnosis remains difficult where laboratory testing is unavailable. This study developed and tested deep-learning models to classify maize nitrogen status from leaf images and evaluated their robustness under real farm conditions. Maize was grown under five nitrogen levels (0, 30, 60, 90 and 130 kg N ha⁻¹) in a field experiment (randomised complete block design) and a pot experiment (completely randomised design) at Sokoine University of Agriculture, and leaf images were captured with a smartphone. Three compact architectures (ResNet34, MobileNetV3-Large and EfficientNet-B0) were trained using five-fold cross-validation on a de-duplicated dataset (3,241 images) and assessed with standard and ordinal metrics, with Grad-CAM for interpretability. All three models were externally validated using independent images collected from 303 smallholder farms in Ruvuma, with nitrogen status labelled based on farmers’ assessments and the Maize Leaf Colour Chart (MLCC). ResNet34 achieved the highest accuracy (93.3%; macro-F1 0.93), followed by MobileNetV3-Large (92.7%) and EfficientNet-B0 (91.2%). Ordinal analysis showed errors concentrated between adjacent classes and Grad-CAM confirmed leaf-focused predictions. Under external validation, exact-classification accuracy declined to 58–64% against MLCC labels and 37–40% against farmer assessment; however, ordinal agreement stayed high (quadratic weighted Kappa 0.86 on MLCC and 0.75 on farmer labels), preserving the correct nitrogen order despite domain shift from genotype, environment, growth-stage, lighting and symptom variability. Smartphone-based models reliably detect nitrogen variation under controlled conditions and, despite lower field accuracy, support rapid nitrogen screening, especially severe deficiency, when combined with farmer knowledge and broader training data. Not applicable.