Journal of Agronomy and Field Innovations

Background: Nutrient deficiencies in agricultural crops result in significant yield loss around the world. Traditional diagnostic approaches either rely on the observation of agronomists or tissue analysis in laboratories, approaches that are not only slow and subjective, but also are difficult to scale. The aim of this study is to introduce an automated framework based on the deep-learning neural network (DNN) which will allow for detecting nutrient deficiency symptoms from leaf images quickly, objectively, and at scale. It will not be the only novelty of this research. 
Methods: The authors used a CNN backbone, which was compared to transformers. The system also included preprocessing (resizing, normalization, contrast improvement, and removal of the background) and augmentation strategies to differentiate between healthy leaves and the leaves facing deficiencies of nitrogen, phosphorus, potassium, magnesium, or iron.
Result: The developed model combining CNN and ViT has reached the maximum possible levels of classification accuracy (96.8%) and F1-score (0.96) in comparison to the architectures evaluated, showing high tolerance towards different lighting conditions and noise. 
Conclusion: The study shows that using deep learning to determine nutrient deficiency allows farmers to add various components to their precision farming techniques and offers scalable and interpretable alternatives to visual identification.