Publicação

Using LUCAS survey and Recurrent Neural Networks to produce LCLU classification based on a Satellite Image time series of Sentinel-2

Ver documento

Detalhes bibliográficos
Resumo:The need of timely and accurate information for the territory has increased over the years, making Land Cover Land Use (LCLU) mapping one of the most common application of remote sensing. Recently, the advances in satellite technology and the open access policies for remote sensing data increased the interest in exploring satellite image time series. In addition, the attention of researchers has shifted from standard machine learning algorithms (e.g., Support Vector Machines and Random Forest) to Recurrent Neural Networks due to their ability of exploiting sequential information. However, acquiring reference data to train these algorithms is still a hurdle. This study aims to evaluate the capability of a Gated Recurrent Unit in performing pixel-level LCLU classification of a satellite image time series, using Sentinel-2 imagery and having the LUCAS survey as reference data. To assess the performance of our model we compared it to state-of-the-art classifiers (SVM and RF). Due to the unbalance nature of the LUCAS survey, we applied oversampling to this dataset to increase the performance of our models, testing three different oversampling techniques. The results attained showed that Recurrent Neural Networks did not outperform the other state-of-the-art algorithms, when trained with a limited number of sampling units, and that oversampling the LUCAS survey increased the performance of all the classifiers. Finally, we were able to demonstrate that it is possible to produce LCLU classification of satellite image time series using only open-source data by using Sentinel-2 imagery and the LUCAS survey as refence data.
Autores principais:Silva, Nuno Alexandre Pereira da
Assunto:LCLU classification LUCAS survey Recurrent Neural Networks Oversampling Sentinel-2
Ano:2021
País:Portugal
Tipo de documento:dissertação de mestrado
Tipo de acesso:acesso aberto
Instituição associada:Universidade Nova de Lisboa
Idioma:inglês
Origem:Repositório Institucional da UNL

Registos relacionados

Registos relacionados