AVANÇOS EM APRENDIZAGEM PROFUNDA APLICADA AO SENSORIAMENTO REMOTO

• Autor(a) principal: Lucas Prado Osco
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFMS
• Texto Completo: https://repositorio.ufms.br/handle/123456789/6720
• topic: Sensoriamento remoto

This thesis consists of an analysis of recent innovations in deep learning (DL) techniques, applied to remote sensing images, with a focus on advancements in Deep Neural Networks (DNN), Visual Language Models (VLM), and zero-shot segmentation with the Segment Anything Model (SAM). The contribution of this work lies in providing a discussion of the state of the art of these technologies within the context of information extraction from remote sensing images. Drawing on literature reviews, model analyses and adaptations, and experiments with remote sensing datasets, the thesis is organized into chapters. The first chapter offers a literature review of the application of DNNs to high spatial resolution images, obtained by sensors onboard Unmanned Aerial Vehicles (UAVs). Here, we analyze 232 scientific articles and demonstrate that DL shows promising results for a range of applications concerning aerial image processing tasks. The second chapter explores the application of Visual ChatGPT, an innovation in VLM, within the remote sensing context. Despite being in the early stages of development, Visual ChatGPT, with its ability to analyze images based on textual inputs, could revolutionize the digital processing of remote sensing images, creating opportunities and optimizing the information extraction process. The third and final chapter examines the performance of SAM in segmenting remote sensing images across multiple scales, representative of varied and challenging geographical contexts. Despite its limitations in images with metric resolution, SAM demonstrates satisfactory performance in segmentation when compared to human manual annotation in multiple cases. In summary, this thesis compiles the latest in the context of applying DL models to multiscale remote sensing images. It establishes both the advancements and challenges to be overcome in this field, outlining paths for future research aimed at assessing remote sensing images in various applications.