Monitoramento dinâmico do volume hídrico em reservatórios utilizando imagens de satélite e redes neurais convolucionais

• Ano de defesa: 2024
• Autor(a) principal: Teixeira, Ariane Marina de Albuquerque
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Sensoriamento remoto - Satélites; Simulação de reservatórios; Monitoramento de reservatórios; Aprendizado profundo; Remote sensing - Satellites; Reservoir simulation; Reservoir monitoring; Deep learning; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/31995

Image-based monitoring techniques have been applied to track water volume. The main objective of this work is to present a framework that leverages deep learning and multispectral images to monitor water volumes in reservoirs in Paraíba, Brazil. To validate the methodology, six different reservoirs throughout the state were selected: Engenheiro Ávidos, Argemiro de Figueiredo, Lagoa do Arroz, Gramame/Mamuaba, Sumé, and Marés. A Convolutional Neural Network (CNN) was employed to extract the surface area of water bodies using satellite images. The volume monitoring graph was derived from the reservoirs' water surface area data, along with their respective elevation-area-volume (EAV) curves. Multispectral satellite images covering the period from mid-2018 to 2023 were employed. Four distinct approaches were tested in creating the segmentation model, with all tests conducted using 5-fold cross-validation. The U-Net-based model, with data augmentation technique, achieved the best metrics, with results of 95.81% IoU and 96.86% Dice coefficient. The temporal series of water volumes highlighted significant variations in the segmentation model performance among the reservoirs, with some, such as Argemiro de Figueiredo and Engenheiro Ávidos, revealing MAPE results below 15% and Pearson coefficients greater than 0.92. In contrast, others, such as Marés and Gramame/Mamuaba, showed inferior metrics. The analysis allowed for evaluating the effectiveness of the model and methodology, identifying factors that may affect the accuracy of estimates, such as cloud presence, spatial resolution of satellite images, and geographical location of the reservoirs.