Identificação de Áreas infestadas por Capim-Annoni por meio de análise espectral

• Ano de defesa: 2025
• Autor(a) principal: Silveira, André Rosa
• Orientador(a): Camargo, Sandro da Silva
• Banca de defesa: Camargo, Sandro da Silva, Neves, Marcos Correa, Breunig, Fábio Marcelo, Moreira, Eduardo Ceretta
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Pampa
• Programa de Pós-Graduação: Mestrado Acadêmico em Computação Aplicada
• Departamento: Campus Bagé
• País: Brasil
• Palavras-chave em Português: Eragrostis plana Nees; Sensoriamento Remoto; Aprendizado de máquina; Validação cruzada; Oversampling; Remote sensing; Machine learning; Cross-validation
• Área do conhecimento CNPq: ENGENHARIAS

The invasion of exotic species represents one of the main threats to biodiversity. In Rio Grande do Sul, *capim-annoni*, introduced in the 1950s, currently occupies about 20% of the grassland vegetation, standing out for its high adaptability and the absence of natural enemies. Therefore, prevention and control measures are essential to mitigate its impacts. This study aimed to apply and compare machine learning algorithms for spectral classification of plant samples, focusing on the identification of capim-annoni (Eragrostis plana Nees) in natural pastures of the Pampa Biome. Reflectance data were collected using an ASD FieldSpec 4 spectroradiometer and, after preprocessing, three datasets were generated: the full set with 1.688 bands, one with 30 bands selected by Random Forest, and another with 28 bands described in the literature. Random Forest, RPart, Support Vector Machines, and Artificial Neural Networks were applied with Leave-One-Out Cross Validation. To address sample imbalance, ADASYN and Mixup were employed. RPart achieved the best performance on the 30-band set (balanced accuracy of 0.727 and sensitivity of 0.800) with ADASYN, followed by Random Forest. These results indicate that spectral analysis has potential to discriminate capim-annoni from native species, although models still favor the majority class, highlighting the need for further studies to improve the identification of the Annoni class.