Deep Learning-Enabled Mobile Application for On-Site Nitrogen Prediction in Strawberry Cultivation

PURPOSE: Precisely applying nitrogen to plants is important for their optimal growth, preventing overuse that may cause water pollution, soil degradation, and financial losses. However, current methods for measuring nitrogen levels are labor-intensive, costly, and destructive, requiring skilled personnel and specialized equipment. METHODS: Thus, in the present study, a novel deep-learning-assisted Android mobile application was developed to predict on-site nitrogen levels in strawberry plants in a non-destructive manner through leaf images. The application categorizes nitrogen levels into eight distinct classes, each representing a 0.25% incremental rise within 2.25% and 4.0% limits. An image dataset was generated by converting chlorophyll readings into nitrogen readings through a linear equation (R² = 0.89). Multiple convolutional neural networks (CNNs) models integrated with residual connections, squeeze and excitation module, convolutional-based attention module (CBAM), and Depthwise convolution were employed to enhance classification accuracy. RESULTS: The CNN model consisting of residual connections and CBAM performed the best with precision, recall, accuracy, and F1-score values of 82.0%, 83.0%, 82.5%, and 82.7%, respectively. In comparison, the proposed CNN, devoid of attention modules, showed the lowest classification accuracy (76.7%). The state-of-the-art models, except MobileNetV1, surpassed the proposed models by up to 3.3% (for EfficientNetB2) in accuracy. CONCLUSIONS: The developed mobile app achieved precision, recall, and accuracy rates of 81.0%, 81.0%, and 80.7%, respectively, offering an efficient, non-destructive method for growers to predict nitrogen levels in strawberry plants, aiding in optimized nitrogen application for sustainable agriculture practices.