Mapeamento de rios em imagens de alta resolução espacial usando algoritmos de aprendizagem de máquina

• Ano de defesa: 2022
• Autor(a) principal: Souza, Mariany Kerriany Gonçalves de
• Orientador(a): Ramos, Ana Paula Marques
• Banca de defesa: Marcato Junior, José , Alves, Marcelo Rodrigo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade do Oeste Paulista
• Programa de Pós-Graduação: Mestrado em Meio Ambiente e Desenvolvimento Regional
• Departamento: Mestrado em Meio Ambiente e Desenvolvimento Regional
• País: Brasil
• Palavras-chave em Português: Aprendizagem de máquina
• Palavras-chave em Inglês: Support Vector Machine; Random Forest; Machine Learning
• Área do conhecimento CNPq: OUTROS::CIENCIAS
• Link de acesso: http://bdtd.unoeste.br:8080/jspui/handle/jspui/1528

Water protection is a problem of interest to several environmental agencies within Pontal do Paranapanema. In this study, an approach for classifying and segmenting rivers in high spatial resolution images based on shallow machine learning algorithms was carried out. Composed of four steps: (1) vectorization of rivers in high spatial resolution images; (2) image segmentation and training of shallow machine learning algorithms; (3) testing the algorithms in the segmentation of the rivers in the RGB images; (4) qualitative description of water courses. The case study area is the region belonging to the 22nd Water Resources Management Unit of Pontal do Paranapanema (UGRHI-22), which includes 26 municipalities in the west of the state of São Paulo. The images used are RGB orthophotographs of high spatial resolution (1 m) from the Mapeia-SP project. To compose the training and test data, a geographic database was organized in GIS ESRI ArcGIS Pro 2.8. Next, we made the classification of interest (rivers x non-river) in these images. A total of 04 orthophotos composed the initial case study. The performance of the Random Forest algorithms was verified, with different evaluation metrics, and the Support Vector Machine (SVM), both applied on RGB images segmented by the Mean Shift method. The performance results, analyzed from the validation metrics of accuracy (90%), such as F1-Score (91%), Recall (99%) and Precision (85%), showed that the SVM presents the best values. Metrics above 90% were obtained in the test data in terms of F1-measure. The Random Forest (RF) input parameters were also tested and showed that the highest regularization parameter obtained a result of 90%. These results were obtained from only 04 images, which demonstrate the ability of shallow algorithms, such as SVM, to perform the mapping and segmentation of rivers in images with a high level of detail. The cartographic products that can be generated from the use of these methodologies based on artificial intelligence can support several studies regarding environmental planning and monitoring tasks and environmental impact analysis related to the protection of water courses.