FORESTRYVIEW APP: MAPEAMENTO DA SILVICULTURA COM USO DE SENSORES REMOTOS, DEEP LEARNING E COMPUTAÇÃO EM NUVEM

• Ano de defesa: 2021
• Autor(a) principal: Joao Otavio Nascimento Firigato
• Orientador(a): Vitor Matheus Bacani
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Google earth engine, segmentação semântica, uso e cobertura da terra, redes neurais convolucionais.
• Link de acesso: https://repositorio.ufms.br/handle/123456789/3856

Mapping land use and land cover is an important tool for monitoring changes in geographic space. From this mapping, it is possi ble to extract information that makes it possible to understand the landscape evolution of a certain area or region. The state of the art in digital classification of satellite images are deep learning algorithms based on convolutional neural networks that were supported by the emergence of cloud computing platforms, where it is possible to obtain a large amount of geospatial and high power data processes used in the training process of these algorithms . This work aimed to map silviculture areas in the east ern mesoregion of the state of Mato Grosso do Sul, using deep learning models for semantic segmentation, making the results of this mapping available through an application accessible on the internet. The methodological procedures were based on the use of a database with 1320 images from the Sentinel 2A satellite MSI sensor for the months of May, July and September 2019 in order to obtain a greater spectral variation of the targets during the year, considering the forestry and non forestry classes. 960 samp les were used for training, 180 samples for validation and 180 for testing. These data feed the neural network, extracting the characteristics of the training images, in order to generate a model that enables the prediction of labels for the rest of the st udy region. The results obtained for the final model showed an overall accuracy of 98% for the test data, making it possible to predict the mapping of the entire study area for the years 2017, 2018, 2019 and 2020. For the availability of the mappings, a Th e application was created on the Google Earth Engine platform, allowing accessibility and interactivity with information from spectral indexes such as NDVI, SAVI, EVI, among others. Thus, the combination of cloud computing, deep learning and remote sensing is highly promising for obtaining geospatial products and analyzes, which are essential in the context of the environmental and socioeconomic planning of the study area.