Machine learning and deep learning applied to computer vision for nutritional status classification in maize (Zea mays L.)

• Ano de defesa: 2025
• Autor(a) principal: Silva, Thiago Lima da
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Atributos de textura; Convolutional neural network; Digital phenotyping; Fenotipagem digital; Nitrogen; Nitrogênio; Rede neural convolucional; Texture attributes
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/11/11152/tde-02122025-093111/

The use of machine learning and deep learning techniques has shown promise for pattern recognition in agricultural images, particularly in the assessment of crop nutritional status. This study aimed to classify maize (Zea mays L.) leaves subjected to different nitrogen doses at two phenological stages, V4 and R1, through the extraction of texture attributes and the use of different machine learning algorithms and convolutional neural network architectures. The leaves were scanned at high resolution, segmented into 224×224-pixel image blocks, and organized into classes corresponding to nitrogen doses: D1 (omission), D2 (half of the recommended dose), D3 (recommended dose), and D4 (excess). For classical methods, the Gray-Level Co-occurrence Matrix was applied to extract attributes such as contrast, correlation, energy, homogeneity, dissimilarity, entropy, mean, and variance, which were later used in classifiers such as Support Vector Machine and Artificial Neural Network. For deep learning-based methods, the architectures AlexNet, DenseNet201, EfficientNetB0, GoogLeNet, NasNetMobile, ResNet50, ResNet101, and VGG19 were evaluated using k-fold cross-validation (k = 10). The results showed that, at the V4 stage, class separation presented greater overlap, leading to moderate performance, especially for AlexNet and GoogLeNet. At the R1 stage, a significant improvement in accuracy and the Kappa coefficient was observed, with deeper architectures such as ResNet50, ResNet101, and DenseNet consistently outperforming classical machine learning methods. It was concluded that deeper networks demonstrate a greater capacity for generalization and discrimination of spectral patterns related to nitrogen in maize, particularly in reproductive stages. Thus, this work contributes to the advancement of digital phenotyping and plant nutritional diagnosis, reinforcing the potential of computer vision and artificial intelligence in agricultural systems.