Abstract

Remote sensing for Earth observation is a growing scientific field essential for many human activities. Among the different applications in the Remote Sensing domain, the production of thematic maps, such as Land Cover and Land Use, are among the most relevant, as this information plays a critical role in management, planning and monitoring activities at different levels. In this context, the Sentinel-2 satellites are of great importance since they provide open data on land and coastal areas at different spatial resolutions (10, 20, and 60 m), democratizing usability, and creating a high potential for the generation of valuable information, useful in many scenarios, such as agriculture, forestry, land cover and urban planning, among others.

 

In this thesis, we aim to exploit the usability of Sentinel-2 data by applying deep learning techniques, which are revolutionizing the world of computer vision and, recently, remote sensing. First, we propose super-resolution models to improve the spatial details of the different Sentinel-2 bands, and second, we propose the conjunction of semantic segmentation with super-resolution to generate improved land cover maps that benefit from the enhanced spatial details of the bands.

 

We first address super-resolution by proposing two different models, one for the 10 m/pixel bands to reach 2 m/pixel and another for the 20 and 60 m/pixel bands to achieve 10 m/pixel. Then, we propose two different multitasking models to derive land cover maps. The first one extending a semantic segmentation model to produce an additional super-resolution image and the second, improving our first super-resolution approach, to provide a semantic segmentation map, as well. We combine features of the different tasks within a single model to improve performance and to generate a high-resolution image with the corresponding highquality land cover map. All models developed were evaluated, quantitatively and qualitatively, using different datasets, showing excellent performance in diverse complex scenarios.