Remote sensing and Deep Learning applied to vegetation mapping

The fast development of human civilization imposed a recent and big environmental impact on the planet earth and the collective of life on Earth to support. The dawn of civilization is a very recent impact on the geologic time scale of the planet. Human needs to have a dimension of the effect of the...

Descripción completa

Detalles Bibliográficos
Autor: Martins, José Augusto Correa
Tipo de recurso: tesis doctoral
Estado:Versión publicada
Fecha de publicación:2023
País:Brasil
Institución:Universidade Federal de Mato Grosso do Sul (UFMS)
Repositorio:Repositório Institucional da UFMS
Idioma:portugués
OAI Identifier:oai:repositorio.ufms.br:123456789/5584
Acceso en línea:https://repositorio.ufms.br/handle/123456789/5584
Access Level:acceso abierto
Palabra clave:Visão Computacional
Sensoriamento Remoto
Aprendizado Profundo
Descripción
Sumario:The fast development of human civilization imposed a recent and big environmental impact on the planet earth and the collective of life on Earth to support. The dawn of civilization is a very recent impact on the geologic time scale of the planet. Human needs to have a dimension of the effect of their actions inside the areas where it lives and in other natural environments to know their environmental impacts and consequences. The will to give an objective answer for this topic guided this research work. This doctoral dissertation presents the results of three years of research as a Doctoral student in the Environmental Technologies program at UFMS (Federal University of Mato Grosso do Sul). During my research, remote sensing and deep learning were the leading scientific fields I studied. The applications of the conjunction of these sciences to analyze the vegetation composition of urban and natural environments in the form of wetlands. We achieved exciting results in applying these techniques, i.e., an F1-score of 91% and an IoU of 73% for urban vegetation segmentation. Moreover, achieving a maximum 97% of F1-score for a specific plant species and 88% average for the whole dataset of 11 wetland plant species. The advances of the experiments conducted during the Doctorate program comprehend a broad range of sensors, from Unmanned aerial vehicles (UAV) that produce centimeter-level data to sensors capable of producing large earth mosaics. We also worked with a broad range of Deep Learning techniques to develop vegetation models. This research work and development can technologically assist the community in improving the understanding of the natural environment that we live in, leading to more resilient, sustainable, and healthy earth environmental systems