2026-02-032026-02-032025-03-28CORRÊA, Alan Breno Soares. Modelos de detecção de nuvens usando redes neurais totalmente convolucionais em imagens multiespectrais do sentinel-2. Orientador: Fabrício José Brito Barros. 2025. 78 f. Dissertação (Mestrado em Engenharia Elétrica) - Instituto de Tecnologia, Universidade Federal do Pará, Belém, 2025. Disponível em: https://repositorio.ufpa.br/handle/2011/17942. Acesso em:.https://repositorio.ufpa.br/handle/2011/17942Accurate cloud detection in satellite imagery is essential for various remote sensing applications, such as environmental monitoring and land cover change analysis. Satellites like Sentinel- 2 play a key role in this context, providing high-resolution imagery at a global scale with a short revisit period (5 days). However, the presence of clouds and cloud shadows poses a major challenge in the preprocessing of these images, hindering the precise extraction of information. Several approaches based on spectral thresholds and deep learning have been developed to mitigate this issue, but there is still room for improvement. This work proposes the use of Fully Convolutional Neural Networks (FCNNs) for cloud segmentation in Sentinel-2 images, exploring different processing levels (L1C and L2A) and combinations of input spectral bands (all bands and RGB+NIR). Models based on the UNet architecture were trained using EfficientNet-B1 and MobileNet-V2 encoders, aiming to compare performance, segmentation efficiency, and the impact of the number of bands. The CloudSen12 dataset, consisting of 10,000 images of 512×512 pixels from different regions around the world and covering diverse atmospheric conditions, was used for the experiments. Quantitative evaluation included metrics such as Accuracy, Intersection over Union (IoU), and F1-Score, while qualitative analysis was performed through visual inspection of the segmentation masks. The results showed that the EfficientNet-B1 encoder achieved the best performance, reaching 95.21% accuracy, 82.74% IoU, and 90.56% F1-Score. Additionally, models trained with only the RGB+NIR bands achieved competitive performance, with 94.87% accuracy, 81.38% IoU, and 89.73% F1-Score. The comparison between processing levels indicated that the removal of atmospheric effects in L2A had little influence on segmentation compared to L1C. Finally, the proposed models outperformed traditional approaches and other architectures from the literature, highlighting the potential of FCNNs to enhance cloud detection in remote sensing applications.ptAcesso AbertoAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Detecção de NuvemSegmentação de imagensSentinel-2Redes neurais totalmente convolucionaisSensoriamento remotoCloud detectionImage segmentationFully convolutional neural networksRemote sensingDissertaçãoCNPQ::ENGENHARIAS::ENGENHARIA ELETRICAPROCESSAMENTO DIGITAL DE SINAISTELECOMUNICAÇÕES