Normalização Adaptativa e Aprendizado Híbrido: Técnicas de Adaptação de Domínio para Análise de Vegetação Transgeográfica

• Ano de defesa: 2025
• Autor(a) principal: MAX HIROITO TIETI
• Orientador(a): Jonathan de Andrade Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Adaptação de domínio; Redes Neurais Profundas; Imagens aéreas; Ortofotos; Aprendizado por transferência.
• Link de acesso: https://repositorio.ufms.br/handle/123456789/12696

This research explores the ability of machine learning models (specifically deep neural networks) to generalize across different contexts and domains. It addresses domain adaptation, i.e., how to ensure a model trained on a specific dataset can perform well on another dataset with different characteristics. This is explored through adaptive normalization and hybrid learning; for this purpose, domain adaptation is investigated by evaluating the performance of pre-trained models on different datasets from distinct ecosystems. For vegetation analysis, the application of deep learning models to satellite/aerial image observation is addressed, which is fundamental for understanding and monitoring vegetation at different scales. The importance of generalization is justified by the model's ability to perform well in different contexts and domains, a crucial factor for the practical application of machine learning models in areas such as ecological and urban planning, where data can vary significantly from one location to another. Data from two Latin American universities with distinct ecosystems were used: UCDB, in Brazil, and UPS, in Ecuador. Preliminary results indicate that, although all tested models (FCN, DeepLabV3, and SegFormer) performed well, FCN stood out in generalization between datasets, with high F1-score, IoU, and accuracy metrics. The study points to areas for improvement, such as normalization adaptation, network enhancement, and parameter fine-tuning. In the future, more segmentation classes, LiDAR data, and datasets from different hemispheres will be incorporated to deepen the analysis of model generalization. With this, we intend to propose techniques for better adaptation of deep learning models to different domains, through semantic segmentation of images, using deep neural networks for the analysis of vegetation and ecosystem images, resulting in direct applications for research in environmental monitoring, precision agriculture, and urban planning, for example.