A carregar...
Publicação
Metodologia SIG para Previsão da Qualidade da Cortiça com Machine Learning
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.36 MB | Adobe PDF |
Orientador(es)
Resumo(s)
O sobreiro (Quercus suber) é uma espécie fundamental para a sustentabilidade ecológica e económica de Portugal, enfrentando, todavia, ameaças crescentes devido a problemas fitossanitários, que em conjugação com os efeitos das alterações climáticas, conduzem ao declínio da espécie. Esta dissertação propõe uma metodologia baseada em Sistemas de Informação Geográfica (SIG) e técnicas de Machine Learning para a previsão do estado fitossanitário e da qualidade potencial da cortiça. O estudo incidiu sobre uma área no concelho da Chamusca, onde foram recolhidos dados de inventário de 137 sobreiros, integrando variáveis dendrométricas medidas in situ com dados provenientes de deteção remota e variáveis biofísicas do terreno. A metodologia envolveu a utilização de dados LiDAR (Light Detection and Ranging) aerotransportados da Direção-Geral do Território para a caracterização estrutural e imagens multiespectrais do satélite Sentinel-2 para o cálculo de diversos índices de vegetação. Foram testadas duas abordagens de modelação na plataforma KNIME, o modelo completo, que integrou todas as variáveis, e o modelo de satélite, focado exclusivamente em variáveis espetrais e as biofísicas do terreno. Para a classificação, aplicaram-se os algoritmos Random Forest (RF), Support Vector Machine (SVM) e o Multilayer Perceptron (RProp MLP). Os resultados revelaram que, no Modelo completo, o SVM apresentou o melhor desempenho com uma exatidão de 81%. Notavelmente, no Modelo Satélite, o algoritmo RProp MLP superou os restantes, alcançando uma exatidão 83% e um índice Kappa de 67%. A análise demonstrou que os índices espectrais, nomeadamente o CI, CVI, LAI e GNDVI, são os principais determinantes do diagnóstico fitossanitário, enquanto as métricas dendrométricas individuais têm um contributo marginal. O estudo mostrou que a integração de tecnologias de Sistemas de Informação Geográfica (SIG) e algoritmos de Machine Learning para o diagnóstico fitossanitário não invasivo dos montados de sobreiro, permite explicar, de forma robusta, o estado fitossanitário das árvores com base em índices de vegetação obtidos a partir de imagens de satélite Sentinel-2, reduzindo a necessidade de realização de inventário florestal tradicional.
The cork oak (Quercus suber) is a fundamental species for the ecological and economic sustainability of Portugal. However, it faces increasing threats due to phytosanitary problems which, together with the effects of climate change, lead to the decline of the species. This dissertation proposes a methodology based on Geographic Information Systems (GIS) and Machine Learning techniques for predicting phytosanitary status and potential cork quality. The study focused on an area in the municipality of Chamusca, where inventory data were collected from 137 cork oak trees, integrating dendrometric variables measured in situ with data obtained from remote sensing and terrain biophysical variables. The methodology involved the use of airborne LiDAR (Light Detection and Ranging) data from the Portuguese Directorate-General for Territory (Direção-Geral do Território) for structural characterization and multispectral Sentinel-2 satellite imagery for the calculation of several vegetation indices. Two modelling approaches were tested in the KNIME platform: the Full Model, which integrated all variables, and the Satellite Model, focused exclusively on spectral and terrain biophysical variables. For the classification, the Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (RProp MLP) algorithms were applied. The results revealed that, in the Full Model, the SVM achieved the best performance, with an accuracy of 81%. Notably, in the Satellite Model, the RProp MLP algorithm outperformed the others, achieving an accuracy of 83% and a Kappa index of 67%. The analysis demonstrated that the spectral indices, namely CI, CVI, LAI, and GNDVI, are the main determinants of phytosanitary diagnosis, while the individual dendrometric metrics have only a marginal contribution. The study showed that the integration of Geographic Information Systems (GIS) technologies and Machine Learning algorithms for the non-invasive phytosanitary diagnosis of cork oak stands makes it possible to explain, in a robust manner, the phytosanitary status of trees based on vegetation indices obtained from Sentinel-2 satellite imagery, reducing the need for carrying out traditional forest inventories.
The cork oak (Quercus suber) is a fundamental species for the ecological and economic sustainability of Portugal. However, it faces increasing threats due to phytosanitary problems which, together with the effects of climate change, lead to the decline of the species. This dissertation proposes a methodology based on Geographic Information Systems (GIS) and Machine Learning techniques for predicting phytosanitary status and potential cork quality. The study focused on an area in the municipality of Chamusca, where inventory data were collected from 137 cork oak trees, integrating dendrometric variables measured in situ with data obtained from remote sensing and terrain biophysical variables. The methodology involved the use of airborne LiDAR (Light Detection and Ranging) data from the Portuguese Directorate-General for Territory (Direção-Geral do Território) for structural characterization and multispectral Sentinel-2 satellite imagery for the calculation of several vegetation indices. Two modelling approaches were tested in the KNIME platform: the Full Model, which integrated all variables, and the Satellite Model, focused exclusively on spectral and terrain biophysical variables. For the classification, the Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (RProp MLP) algorithms were applied. The results revealed that, in the Full Model, the SVM achieved the best performance, with an accuracy of 81%. Notably, in the Satellite Model, the RProp MLP algorithm outperformed the others, achieving an accuracy of 83% and a Kappa index of 67%. The analysis demonstrated that the spectral indices, namely CI, CVI, LAI, and GNDVI, are the main determinants of phytosanitary diagnosis, while the individual dendrometric metrics have only a marginal contribution. The study showed that the integration of Geographic Information Systems (GIS) technologies and Machine Learning algorithms for the non-invasive phytosanitary diagnosis of cork oak stands makes it possible to explain, in a robust manner, the phytosanitary status of trees based on vegetation indices obtained from Sentinel-2 satellite imagery, reducing the need for carrying out traditional forest inventories.
Descrição
Dissertation presented as the partial requirement for obtaining a Master's degree in Geographic Information Systems and Science, specialization in Geographic Information Systems and Science
Palavras-chave
Quercus suber Machine Learning SIG Deteção Remota Índices de Vegetação Estado Fitossanitário GIS Remote Sensing Vegetation Indices Phytosanitary Status