AVALIAÇÃO E GERAÇÃO DE MODELO DIGITAL DE TERRENO COMBINANDO FOTOGRAMETRIA E VISÃO COMPUTACIONAL

Abstract

Digital Terrain Models (DTMs) are an important source of information for studies on flooding, flash floods, and surface runoff dynamics, as they help to understand the flow paths of water. Despite the wide availability of global Digital Surface Models (DSMs) and the advances in the use of Unmanned Aerial Vehicles (UAVs) for 3D reconstruction, few studies have systematically evaluated the vertical accuracy of these products in urban areas and, more importantly, proposed automated strategies for removing above-ground elements such as vegetation and buildings. This thesis aims to fill this gap by integrating traditional photogrammetric methods with deep learning techniques to enable the direct generation of urban DTMs. In the first chapter, a vertical accuracy assessment was conducted for open-access global DSMs, as well as the commercial TanDEM-X model, in the urban area of Campo Grande, MS, Brazil. Ground control data with centimeter-level planialtimetric accuracy, acquired via GNSS, were used as a reference to evaluate the suitability of these models at different map scales according to the Cartographic Accuracy Standard for Digital Cartographic Products (PEC-PCD). Results indicated that both the TanDEM-X and ALOS AW3D30 models meet the requirements for 1:50,000 mapping, class A of PEC-PCD. In the second chapter, a DSM was generated from UAV imagery, and the influence of the number and distribution of Ground Control Points (GCPs) on the accuracy of the model was analyzed. Different numbers of GCPs (2 to 6) were tested, along with two configurations using six GCPs: one with all points on a flat surface and another including elevated points. It was found that including GCPs in elevated positions significantly improved the quality of the DSM. In the third chapter, five deep learning architectures (FCN, GCNet, PointRend, PSANet, and SegFormer) were applied for tree segmentation in UAV imagery, aiming to directly generate DTMs from SfM-MVS data using only points classified as ground in the mask regions. The resulting DTMs were compared with those obtained using the traditional Cloth Simulation Filtering (CSF) method. Results showed that the DL-based approach yields DTMs with accuracy comparable to CSF, while offering greater potential for automation and scalability. It is concluded that the integration of UAV data with semantic segmentation through deep learning represents a viable and innovative approach for efficient urban DTM generation, with practical applications in hydrological modeling, urban planning, and environmental studies.