Sensoriamento remoto multiespectral para o monitoramento da cobertura vegetal em estruturas geotécnicasSensoriamento remoto multiespectral para o monitoramento da cobertura vegetal em estruturas geotécnicas

Abstract

Geodynamic events such as the collapse of dams, dikes, and other geotechnical structures lead to severe impacts on the environment, infrastructure, properties, and human lives. Brazil ranks eighth in the world in terms of the number of large dams, highlighting the need for continuous monitoring of these structures. The National Dam Safety Policy (PNSB) was established to assign legal responsibility to the enterprises to maintain safety conditions during the construction, operation, and decommissioning of dams. Remote sensing-based monitoring is an efficient alternative for inspecting large infrastructure such as dams and dikes, presenting itself as an alternative to the time-consuming traditional field methods. Multispectral remote sensing data captured via Unmanned Aerial Vehicle (UAV) enables the acquisition of high-resolution images of the structure, allowing for subsequent analyses related to structural health with the aid of Machine Learning and Computer Vision techniques. This doctoral thesis proposal presents an end-to-end methodology for monitoring the vegetation coverage of embankment dam slopes and dikes. The method comprises the automated collection of multispectral data, data processing to obtain Digital Orthophoto Maps (DOMs), semantic land-cover segmentation of the structures, and vegetation health analysis. The proposed methodology was applied in a case study on the geotechnical structures of the Belo Monte Hydroelectric Plant Complex, located north of the Xingu River, in the southwestern region of the state of Pará, Brazil. The land-cover segmentation results achieved an F1 Score of 96,41% and a mean IoU of 93,31% for vegetation cover segmentation of the studied structures, enabling precise analysis of vegetation health metrics. The extraction of vegetation cover from the slopes allows vegetation health assessment based on multispectral vegetation indices, highlighting areas with varying vegetation densities and exposed soil. Additionally, to improve accessibility, a generative AI model was developed using the Pix2Pix cGAN architecture to generate synthetic vegetation indices from RGB images, eliminating the need for costly multispectral sensors. The Pix2Pix models achieved structural similarity indices of 0,95 and 0,94 for the generation of NDVI and NDRE indices, respectively, resulting in high- quality synthetic images. The proposed methodology aims to add redundancy to decision-making processes related to dam safety, contributing to risk mitigation and accident prevention.