Uma nova abordagem para classificação e avaliação de carbono e nitrogênio de solos por análise de imagem: um caso de estudo com latossolos e cambissolos

• Ano de defesa: 2025
• Autor(a) principal: Rinaldi, Juliano Tiago
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Pato Branco; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e de Computação; UTFPR
• Palavras-chave em Português: Solos; Carbono; Solos - Teor de nitrogênio; Aprendizado do computador; Soils; Carbon; Soils - Nitrogen content; Machine learning; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA; Engenharia/Tecnologia/Gestão
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/38389

The accurate classification of soils and the prediction of their chemical properties, such as carbon and nitrogen content, are fundamental in several fields, including agriculture, geology, and environmental management. Traditional methods for soil classification and analysis are often expensive, time-consuming, and dependent on specialized professionals, which limits their sca- lability, especially in regions with diverse terrain or limited resources. This study proposes two complementary approaches based on image analysis and deep learning. In the first study, focused on soil classification, a pre-trained convolutional neural network (ResNet50) achieved a balanced accuracy of 93.73% on the test set. In contrast, applying a majoritarian voting method to the window predictions from the first experiment resulted in an impressive balanced accuracy of 100% in distinguishing between Latosols and Cambisols, two common soil types in southern Brazil. In the second study, we aimed to approximate the predictions of carbon and nitrogen content to the real measured values. For this, a fine-tuned EfficientNetV2-S model was used for regression, with two experiments conducted: one predicting values for segmented windows from the images and another using the median of the window predictions for each soil sample. The latter approach significantly improved performance, achieving an R² of 0.7425 for carbon and 0.7774 for nitrogen. These methods offer faster, non-destructive, and scalable alternatives to traditional laboratory analyses, contributing to sustainable agricultural practices and facilitating soil classification and management in resource-limited contexts.