A carregar...
Publicação
Identificação da ocupação e uso do solo com base em imagens provenientes de deteção remota e em algoritmos de machine learning: a Reserva da Faia Brava como caso de estudo
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 10.75 MB | Adobe PDF |
Orientador(es)
Resumo(s)
Esta dissertação procura identificar a ocupação e uso do solo com base em imagens de deteção remota e algoritmos de Machine Learning (ML), utilizando como estudo de caso a Reserva da Faia Brava, situada no vale do Côa, distrito da Guarda, Portugal. O estudo avalia a exequibilidade de ferramentas de código aberto, como o QGIS e o plugin Orfeo Toolbox, para implementar fluxos de trabalho de classificação supervisionada. Foram processadas imagens de alta resolução obtidas por UAV (2,7 cm/pixel) e pelo satélite Pléiades-Neo (30 cm/pixel), integrando índices de vegetação (GLI, NDVI, SAVI e MSAVI) e métricas texturais de Haralick. O treino dos modelos foi realizado numa quadrícula com 500 metros de lado, selecionada pela sua diversidade ecológica, e posteriormente testada em outras áreas da reserva para avaliar a capacidade de generalização. Dois algoritmos de ML, Random Forest (RF) e Support Vector Machine (SVM), foram testados, com desempenhos avaliados através de matrizes de confusão, F1-scores e coeficientes Kappa. Os resultados evidenciaram a superioridade dos ortofotomosaicos UAV face às imagens de satélite, especialmente quando combinados com análise textural, embora tenham sido identificadas limitações relacionadas com variações sazonais da vegetação e a interoperabilidade entre sensores. O algoritmo RF mostrou maior consistência enquanto o SVM revelou sensibilidade à complexidade espectral. O estudo destaca a aplicabilidade prática destes métodos para a monitorização ambiental, sublinhando a importância das soluções open source para a democratização das tecnologias de deteção remota. Como produto final foi produzido, com base nos modelos de ML, uma carta temática para a área total da Reserva da Faia Brava, abrangendo quatro classes: árvores e arbustos, vegetação herbácea, afloramentos rochosos e outras ocupações e usos do solo.
This dissertation uses the Faia Brava Reserve, located in the Côa Valley in the Guarda district of Portugal, as a case study to identify land use land cover based on remote sensing images and Machine Learning (ML) algorithms. The study evaluates the feasibility of using open-source tools such as QGIS and the Orfeo Toolbox plugin to implement supervised classification workflows. High-resolution images obtained from a UAV (2.7 cm/pixel) and a Pléiades-Neo satellite (30 cm/pixel) were processed and integrated with vegetation indices (GLI, NDVI, SAVI and MSAVI) and Haralick textural metrics. The models were trained on a 500-metre square area selected for its ecological diversity and then tested in other areas of the reserve to evaluate their generalisation capacity. Two ML algorithms, Random Forest (RF) and Support Vector Machine (SVM), were tested and their performance was assessed using confusion matrices, F1-scores and Kappa coefficients. The results demonstrated the superiority of UAV orthophotomosaics over satellite images, particularly when combined with textural analysis. However, limitations relating to seasonal variations in vegetation and interoperability between sensors were also identified. The RF algorithm demonstrated greater consistency, whereas the SVM was sensitive to spectral complexity. The study highlights the practical applicability of these methods for environmental monitoring and underlines the importance of open-source solutions in democratising remote sensing technologies. The final product, based on ML models, was a thematic map of the entire Faia Brava Reserve area, divided into four categories: trees and shrubs, herbaceous vegetation, rocky outcrops, and other land uses land covers.
This dissertation uses the Faia Brava Reserve, located in the Côa Valley in the Guarda district of Portugal, as a case study to identify land use land cover based on remote sensing images and Machine Learning (ML) algorithms. The study evaluates the feasibility of using open-source tools such as QGIS and the Orfeo Toolbox plugin to implement supervised classification workflows. High-resolution images obtained from a UAV (2.7 cm/pixel) and a Pléiades-Neo satellite (30 cm/pixel) were processed and integrated with vegetation indices (GLI, NDVI, SAVI and MSAVI) and Haralick textural metrics. The models were trained on a 500-metre square area selected for its ecological diversity and then tested in other areas of the reserve to evaluate their generalisation capacity. Two ML algorithms, Random Forest (RF) and Support Vector Machine (SVM), were tested and their performance was assessed using confusion matrices, F1-scores and Kappa coefficients. The results demonstrated the superiority of UAV orthophotomosaics over satellite images, particularly when combined with textural analysis. However, limitations relating to seasonal variations in vegetation and interoperability between sensors were also identified. The RF algorithm demonstrated greater consistency, whereas the SVM was sensitive to spectral complexity. The study highlights the practical applicability of these methods for environmental monitoring and underlines the importance of open-source solutions in democratising remote sensing technologies. The final product, based on ML models, was a thematic map of the entire Faia Brava Reserve area, divided into four categories: trees and shrubs, herbaceous vegetation, rocky outcrops, and other land uses land covers.
Descrição
Palavras-chave
Deteção remota Machine learning Qgis Orfeo toolbox Ocupação e uso do solo Faia Brava