Land Cover mapping based on Hierarchical Decision Trees

• Autor(a) principal: Feio, Clarisse Rita Afonso
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/126158
• topic: remote sensing; land cover classification; Sentinel-2; decision trees; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

The ability to monitor land cover changes can be very useful for resource management, urban planning, forest fire identification, among plenty of other applications. The topic of remote sensing has been studied for a long time, with many different solutions that typically use satellites or aircraft to obtain multi-spectral imagery and further analyse it. The entity responsible for monitoring land use and land cover in Portugal is Direção- Geral do Território (DGT) which periodically produces a document called Land Use and Land Cover Map (Carta de Uso e Ocupação do Solo (COS), in Portuguese). This document uses imagery with high spatial resolution of 0,25 m and has a minimum mapping unit of 1 ha, however, it is only produced every few years because it is manually curated by experts. This hinders the ability to closely monitor relevant land changes that occur more frequently or rapidly. In this dissertation, several classifiers were developed in a hierarchical manner to address some of COS drawbacks. The classifiers used were based on decision trees which were trained using satellite imagery collected from Sentinel-2 satellite constellation. Although having a lower spatial resolution than COS, they can automatically classify land cover in some minutes every time a new set of Sentinel-2 imagery is collected, in this case each 5 days. Cloud coverage might make some of these images unusable but nonetheless, the temporal resolution is still far greater than COS. However, automatic classification is not as accurate as manual classification. The produced classifiers did not consider as many classes as COS and had problems distinguishing some types of land cover, due to either poor sample size or spectral signature similarity. Considering Matthews Correlation Coefficient (MCC), water class had the best performance with an average of 91,28%, followed by forest and agriculture class with an average of 47,88% and 42,34%, respectively, and lastly urban areas and bare land class had the worse results averaging 28,03% and 20,53% respectively. Nevertheless, the results obtained were still considered to be good, but with considerable room for improvement.