Avaliação temporal da detecção de deficiências nutricionais em folhas de café por deep learning

Abstract

The diagnosis of nutritional deficiencies in coffee cultivation, a cornerstone of Brazilian agribusiness, is traditionally reactive and subjective. This work proposes a temporal analysis for the early detection of nutritional stress, with the objective of determining from which point in time symptoms become visually discriminative for deep learning models. To this end, three deep learning architectures were systematically compared, namely ResNet50, a Vision Transformer (ViT-B/16), and a hybrid CNN-ViT model, in the classification of six nutritional deficiencies (N, P, K, Ca, Mg, S) and a control class, using an RGB image dataset collected over five months. The results of 98 experiments demonstrate that the performance of the classifiers is strongly dependent on the symptom stage, with the best observed performance in the configuration that retains the final 40% of the experimental period. Under the standardized protocol adopted, the ResNet50 architecture outperformed the Hybrid CNN-ViT, which in turn outperformed the ViT-B/16, with the convolutional architecture reaching a maximum accuracy of 77.97% with 512 × 512 pixel images, against 71.68% for the Hybrid and 67.74% for the ViT-B/16. An exploratory interpretability analysis via Grad-CAM produced activation maps visually consistent with the expected agronomic symptomatology of each deficiency. Under the controlled experimental conditions, the inductive locality bias of CNNs proved more suitable for capturing foliar texture patterns in this domain, and temporal analysis is crucial for understanding the limits and potential of automated early detection.