Researchers have evaluated an explainable artificial intelligence framework designed to automatically classify nutrient deficiencies in coffee leaves under both controlled laboratory settings and simulated field-like environments. Visual inspection in the field is frequently subjective, while chemical laboratory analyses remain costly and time-consuming for routine monitoring. Using the CoLeaf-DB dataset, the study assessed automated diagnostic approaches capable of identifying eight individual mineral deficiencies, healthy leaves, and instances involving multiple concurrent deficiencies.
To simulate challenging visual conditions encountered during real-world agricultural image acquisition, the authors incorporated leaf segmentation, randomized background generation, and photometric augmentation directly into the model training pipeline. These synthetic augmentations were designed to introduce visual variability, curtail shortcut learning, and improve real-world generalization across distinct machine learning models evaluating coffee foliage.
In comparative evaluations across representative classical and deep learning architectures, a Support Vector Machine demonstrated the highest classification performance. The Support Vector Machine model achieved an accuracy of 74.9 percent on the clean test set and 75.3 percent on the simulated field-like test set, outperforming competing architectures under both test configurations.
Other classical machine learning architectures also maintained competitive diagnostic capability when tested against simulated field variability. A Multi-Layer Perceptron model reached a classification accuracy of 73.7 percent under simulated field-like conditions, while an XGBoost model recorded an accuracy of 73.3 percent on the same augmented test dataset.
Deep learning models showed substantial performance improvements when pre-trained representations were leveraged through transfer learning. For the ResNet50 neural network, transfer learning increased classification accuracy on clean test images from 67.1 percent under scratch training to 71.4 percent. On the simulated field-like dataset, transfer learning boosted ResNet50 accuracy from 62.4 percent to 70.2 percent, marking gains of 4.3 and 7.8 percentage points, respectively.
To verify that model predictions relied on biologically meaningful features rather than background artifacts, the researchers employed interpretability tools including Grad-CAM, LIME, and SHAP. An illustrative analysis conducted on a representative potassium-deficiency leaf confirmed that the model decision was specifically associated with chlorotic and necrotic leaf areas rather than surrounding background elements.
Despite promising results for individual nutrient classes, the framework revealed significant diagnostic limitations when handling simultaneous nutrient shortages. The multiple-deficiency class recorded an F1-score of 0.129, a sensitivity of 0.095, and a Matthews correlation coefficient of 0.087, demonstrating that the current architecture is substantially less reliable when multiple mineral deficiencies occur at the same time.
Subject of Research: Automated classification of coffee leaf nutrient deficiencies using explainable artificial intelligence
Article Title: Explainable Artificial Intelligence Framework for Robust Coffee Leaf Nutrient Deficiency Classification Under Simulated Field-Like Conditions
Article References: Acici, K., Bademci, M., Ekinci, F., Özdemir, M., Gümrükçü, A., Krupodorova, T., Akata, I., Aydın, A., Sevindik, M., & Guzel, M. S. (2026). Explainable Artificial Intelligence Framework for Robust Coffee Leaf Nutrient Deficiency Classification Under Simulated Field-Like Conditions. International Journal of Computational Intelligence Systems. https://doi.org/10.1007/s44196-026-01666-5
Image Credits: AI Generated
DOI: 10.1007/s44196-026-01666-5
Keywords: explainable AI, coffee leaves, nutrient deficiency, CoLeaf-DB, Support Vector Machine, ResNet50, transfer learning, XGBoost, Multi-Layer Perceptron, Grad-CAM, LIME, SHAP, precision agriculture, plant phenotyping, synthetic augmentation, computer vision
Cite Scienmag News
Blake Davidson. (November 19, 2026). Explainable Machine Learning Framework Classifies Coffee Leaf Nutrient Deficiencies Under Simulated Conditions. Scienmag. https://scienmag.com/explainable-machine-learning-framework-classifies-coffee-leaf-nutrient-deficiencies-under-simulated-conditions/
Blake Davidson. "Explainable Machine Learning Framework Classifies Coffee Leaf Nutrient Deficiencies Under Simulated Conditions." Scienmag, 19 November 2026, https://scienmag.com/explainable-machine-learning-framework-classifies-coffee-leaf-nutrient-deficiencies-under-simulated-conditions/. Accessed 19 November 2026.
Blake Davidson. "Explainable Machine Learning Framework Classifies Coffee Leaf Nutrient Deficiencies Under Simulated Conditions." Scienmag. November 19, 2026. https://scienmag.com/explainable-machine-learning-framework-classifies-coffee-leaf-nutrient-deficiencies-under-simulated-conditions/








