Intra-Annual land cover mapping: Automatic training sample extraction from old maps for intra-annual land cover mapping at central of Portugal

Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial TechnologiesMaking operational e cient the production of Land Use Land cover (LULC) mapping over large areas as the consistency and accuracy keep a high quality is an essential condition for the implementation of applications that require periodic information, such as forest re propagation, crop monitoring or climate models. The increasing spatial and temporal resolution satellite images, such as those provided by Sentinel 2, open new opportunities for producing accurate datasets that can improve the lack of production of global and regional LULC maps with ne scale and up-to-date information. In this context, while this thesis aimed to make automatic the generation of intra-annual maps implementing a work ow that consists of supervised classi cation in synergy with automatic extraction of training samples from an old map, it also aimed to use singular and BAP composites. Therefore, after a preliminary selection and preprocessing of the implemented spectral bands in the classi cation both from single and BAP composites of Sentinel 2 images of 2017, a random selection of training points is extracted from an old reference map; national LULC map of Portugal, COS 2015. We performed a classi cation scheme using support vector machine (SVM) and Random forest (RF) classi ers with two datasets of six and nine di erent number of land cover classes. The out-of-date information derived from the old map led us to evaluate the viability of implementing two re ning procedures over the data to improve accuracy; one based on margins of NDVI signals and another based on an iterative learning procedure. Since the proposed methodologies did not lead to improving OA on the classi cation of any of the images of 2017, we questioned for robustness of the classi ers RF and SVM by injecting di erent levels of noise during the modeling. Finally, the free cloud and phenological maximization of the BAP composites become in a consistent and e cient input for the production of seasonal LULC mapping