A carregar...
Publicação
Super-resolução de imagens Sentinel-2 para melhoria da classificação da ocupação do solo utilizando redes generativas adversárias (GAN)
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 5.08 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
The 10-meter spatial resolution of Sentinel-2 images can be a limiting factor in the detailed analysis of land use and land cover. This dissertation examines the application of super-resolution techniques based on Generative Adversarial Networks (GANs) to improve the spatial resolution of Sentinel-2 images and to evaluate their impact on land cover classification. A set of very high-resolution orthophotos (25 cm) was used as reference to train a GAN model capable of producing super-resolved images at 2.5 m resolution, corresponding to a fourfold increase in detail. The evaluation of the results included quantitative image quality metrics (Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Mean Squared Error (MSE), Mean Absolute Error (MAE)) and visual comparisons with the original images. In addition, the influence of the super-resolved images on thematic classification was assessed using a Random Forest classifier, comparing performance in discriminating against urban land cover classes (buildings, roads, vegetation, and water) between the original and super-resolved images. The results demonstrate substantial improvements in image quality: the super-resolved product achieved PSNR values approximately 3 dB higher and SSIM values nearly three times greater than those of the original images, closely approaching the very high-resolution reference. Visually, the results exhibit a higher level of detail and greater clarity in the identification of urban structures. In land cover classification, performance gains were observed with the use of super-resolved images, reflected in increases in F1 values for critical classes such as buildings and roads, bringing classification accuracy closer to that obtained with very high-resolution data. This work highlights the potential of Generative Adversarial Networks to overcome the resolution limitations of freely available remote sensing imagery, contributing to improvements in urban mapping and territorial monitoring.
Descrição
Trabalho de Projeto de Mestrado, Engenharia Geoespacial, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Image super-resolution Sentinel-2 Generative Adversarial Networks land cover classification remote sensing