Recomendações de manejo nutricional e prevenção de estresse abiótico com base na classificação de deficiências em folhas de plantas utilizando visão computacional e deep learning

Abstract

Plants are essential for planet Earth, serving as the primary producers of oxygen, accounting for the storage of over 90% of the world’s terrestrial biomass, and forming the base of the food chain. However, climate change poses challenges to strengthening food security, increasing agricultural productivity, and improving crop quality, which can be adversely affected by extreme events such as droughts and floods. These environmental factors can lead to abiotic stresses, impacting the economy and agricultural production, what can make the constraints on global food security more severe. Proper nutrient management, therefore, emerges as an effective solution, promoting natural defenses and enhancing stress tolerance, as nutritional imbalances in plants affect not only their growth but also their defense mechanisms and tolerance to abiotic stresses. Consequently, the early and effective identification of nutritional deficiencies is crucial to ensure greater efficiency in agriculture. Among the existing diagnostic methods, these deficiencies can be identified through visual symptoms such as chlorosis, necrosis, and deformations along the leaf edges. In this context, this work proposes the development of a system that integrates a CNN model, trained based on the EfficientNetV2-B3 architecture, designed to categorize coffee plant leaves into six distinct classes: healthy leaves and deficiencies of K, Mn, B, Ca, and P. Additionally, the system includes a recommendation pipeline based on the RAG approach, built using ChromaDB, LangChain, and the GPT-4o model, which enables the generation of recommendations to support efficient and environmentally responsible management of nutritional deficiencies identified by the classification model. The final results of the classifier indicated an overall accuracy of 94%, a macro-average (F1) of 92%, and a weighted-everage (F1) of 94%, demonstrating robust performance with potential to assist in rapid and precise nutritional diagnostic processes. Moreover, the management recommendations were evaluated using RAGAS across six distinct metrics, which indicated that, in general, the responses generated by the LLM are relevant, clear, grounded in the information extracted from the vector database, and consistent with the identified nutritional deficiency.