IMPLEMENTAÇÃO E AVALIAÇÃO DE MÉTODOS DE APRENDIZAGEM PROFUNDA PARA MAPEAMENTO DE POSTES EM ORTOIMAGENS AÉREAS

Abstract

Currently, asset management is a process of great importance within an energy concessionaire, since it is fundamental to the decision-making of the corporation and is directly linked to the company’s economic and financial balance. A great ally in this process are georeferenced coordinates for each asset, assisting the corporation in planning new investments and also providing a complete overview of the company’s infrastructure. Glimpsing the possibility of greater regulatory control through geospatial data, as of 2009, ANEEL (National Electric Energy Agency) instituted the Distributor’s Geographic Database (BDGD), which imposed energy concessionaires to annually provide the georeferenced register of your assets. ANEEL does not determine a specific geo-registration methodology for companies, which still often dispense the use of technological resources and opt for costly and archaic working methodologies, mapping assets through professionals who inspect the electricity grid street by street. This dissertation aims to analyze two proposals for deep learning methodologies that assist in the automatic georeferencing of electrical assets through aerial orthoimages. The first proposal is to use the Adaptive Training Sample Selection (ATSS) architecture in this task, a novel method and still not often used in the field of remote sensing. The performance of the method will be compared with the Faster R-CNN and RetinaNet architectures, already commonly used in the field of remote sensing, in the execution of the same task. The second proposal is the implementation and analysis of ATSS, TOOD (Task-aligned One-stage Object Detection), Varifocal-Net and Deformable DETR (Detection Transformer) methods in the detection and classification task of six classes of poles present in the dataset.