Classificação de Regiões em Imagens do Satélite Sentinel 2 através de Inteligência Artificial
Carregando...
Arquivos
Data
2023-02-28
Autores
Galvão, Jonilson Michel Fontes
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A disponibilidade de recursos para estudo e compreensão da dinâmica do uso e cobertura do solo através de sensoriamento remoto por meio de técnicas de classificação com auxílio de aprendizado de máquina tem ganho destaque nos últimos anos. Este trabalho visa apresentar resultados da classificação de regiões em imagens do satélite Sentinel-2 da Área de Preservação Ambiental de Itupararanga, especificamente no entorno da represa de Itupararanga, no Estado de São Paulo. Através do pré-processamento e utilização de índices espectrais (NDVI, GNDVI, NDWI, MNDWI e NDBI) foi feita a classificação da cobertura do solo da região usando as técnicas de inteligência artificial: Random Forest (RF) e Support Vector Machine (SVM). As métricas de avaliação de desempenho registraram excelentes resultados com 99,27% de acurácia global para SVM com kernel radial e 99,15% para RF. O cálculo da cobertura do solo realizado por meio do SVM com kernel radial resultou em 3,4% de agricultura, 7,0% de água, 24,0% de vegetação arbórea, 4,2% de eucalipto, 27,7% de vegetação rasteira, 14,9% de solo exposto, 7,2% de interação solo e arbusto, 14,0% de área urbana e 0,2% de nuvem nas imagens analisadas.
The availability of resources for studying and understanding of land cover dynamics with remote sensing and classification techniques based on machine learning has gained prominence in recent years. This work aims to present results of the classification of regions in Sentinel-2 satellite images from Itupararanga Environmental Preservation Area, specifically around Itupararanga dam, in the State of São Paulo. Through the pre-processing and use of spectral indexes (NDVI, GNDVI, NDWI, MNDWI and NDBI), soil cover of the region was classification using the following artificial intelligence techniques: Random Forest (RF) and Support Vector Machine (SVM). Performance evaluation metrics showed excellent results with 99,27% overall accuracy for SVM with radial kernel and 99,15% for RF. The calculation of land cover performed by means of SVM with radial kernel resulted in 3,4% of agriculture, 7,0% of water, 24,0% of tree vegetation, 4,2% of eucalyptus, 27,7% of undergrowth, 14,9% of exposed soil, 7,2% of soil and shrub interaction, 14,0% of urban area and 0,2% of cloud in the analysed images.
The availability of resources for studying and understanding of land cover dynamics with remote sensing and classification techniques based on machine learning has gained prominence in recent years. This work aims to present results of the classification of regions in Sentinel-2 satellite images from Itupararanga Environmental Preservation Area, specifically around Itupararanga dam, in the State of São Paulo. Through the pre-processing and use of spectral indexes (NDVI, GNDVI, NDWI, MNDWI and NDBI), soil cover of the region was classification using the following artificial intelligence techniques: Random Forest (RF) and Support Vector Machine (SVM). Performance evaluation metrics showed excellent results with 99,27% overall accuracy for SVM with radial kernel and 99,15% for RF. The calculation of land cover performed by means of SVM with radial kernel resulted in 3,4% of agriculture, 7,0% of water, 24,0% of tree vegetation, 4,2% of eucalyptus, 27,7% of undergrowth, 14,9% of exposed soil, 7,2% of soil and shrub interaction, 14,0% of urban area and 0,2% of cloud in the analysed images.
Descrição
Palavras-chave
Geoprocessamento, Aprendizado do computador, Inteligencia artificial