Avaliação da qualidade da segmentação de imagem do CBERS 04A através do modelo U-Net no contexto do planejamento urbano no município de Itaqui - RS

Abstract

The segmentation process is crucial to extracting relevant information from remote sensing images, such as identification of urban areas, plant cover, water bodies and other geographical elements. However, the manual segmentation of satellite images can be laborious and slow, and is prone to human errors. Convolutional neural networks (CNNs) offer an advanced and automated alternative to the manual segmentation of satellite images, allowing for more accurate and efficient analysis of the spatial and spectral characteristics of the images. With this in mind, the present work brings an in-depth analysis of the development and performance of a U-Net convolutive neural network model, able to segment images from the CBERS 04A satellite. The great advantage of using a CNN is its ability to automatically learn complex characteristics of images, providing accurate and efficient results in targeting and remote sensing data analysis tasks. The model chosen was the U-Net that uses convolutionary layers to capture details on multiple scales on input and output, facilitating precise segmentation of images by combining downsampling and upsampling path information. As the training data (masks) still need to be done manually, there is a great chance of there being systematic errors, thanks to the poor quality of the cuts of the image of the urban area of Itaqui. The model is able to correctly segment most pixels, but there is room for improvements, besides better quality images can significantly increase the accuracy of the model. The resulting segmentation can be used in conjunction with other data to assist the urban planning of the city of Itaqui, especially with regard to the orderly growth of the town and prevention in cases of flooding which is a recurring problem in the region.