Identifying nutritional deficiencies at an early stage is crucial for maximizing yield production and ensuring healthy plants. Conventional methods generally rely on time-consuming analysis conducted by agronomic experts. To address this challenge, this study presents a data-driven approach for the early identification of nutritional deficiencies in hazelnut orchards. Different custom datasets, composed of images acquired in a real hazelnut orchard as well as in a controlled laboratory environment, are collected, and the performance of five state-of-the-art machine learning models in early detecting nutritional deficiencies is compared. In particular, ResNet, DenseNet, MobileNet, EfficientNet, and ConvNext models, along with a baseline based on support vector machines, are considered. Data augmentation techniques are introduced to synthetically increase the datasets, and their effectiveness is extensively evaluated. Additionally, a pipeline is designed to carry out the early identification of nutritional deficiencies onboard an agricultural robot. Experimental results on the early identification show that ConvNext achieves the highest performance: 81.79% accuracy and 0.8168 F1 score on a real-world dataset with four classes, and 75.54% accuracy with 0.7552 F1 score for the more challenging six-class scenario. Furthermore, the effectiveness of the integrated system is validated in preliminary laboratory experiments using a Turtlebot2 mobile base and a Franka Research 3 arm, equipped with RGB-D cameras. • Data-driven pipeline detects hazelnut nutrient deficiencies from leaf images. • Real and lab-acquired hazelnut leaf image datasets collected and publicly released. • ConvNext achieves 85.50% accuracy on 4-class hazelnut deficiencies in lab conditions. • 75.54% accuracy on 6-class real orchard dataset validates field robustness. • Two-stage pipeline with leaf detection and classification enables onboard robot monitoring.
Fuoti et al. (Sat,) studied this question.