Enhancing Spatial Resolution in Sentinel-3 Data - A Landsat 8 Supervised Approach

The temperature of Earth's surface has become in recent years critical for various research areas including weather forecasting, environmental monitoring, vegetation health analysis, urban planning, human well-being, and agricultural management. A major issue arising from urbanization is the development of urban heat islands — a phenomenon analyzed within microclimate studies, which focuses on localized areas that exhibit significantly higher temperatures than their surroundings. This challenge has increased the demand for finer spatial resolution of Land Surface Temperature (LST), enabling stakeholders and policymakers to make informed decisions and implement effective strategies within their areas of responsibility. Remote sensing instruments on satellites have been a main tool for estimating LST. In this case study datasets of temperature measurements come from Sentinel-3 and Landsat 8, which are two satellite missions that use spectral bands to calculate LST, each with its own limitations: Sentinel-3 provides multiple readings per day but with a spatial resolution of 1 km, whereas Landsat 8 offers a finer spatial resolution of 100 m but revisits the same area only every eight days. Moreover the two satellites' dataset have different types of data content and these discrepancies are addressed to make them as similar as possible. The goal of this thesis is to prepare data from the aforementioned satellites — along with auxiliary data including Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), Land Cover (LC), and weather data that act as additional information — and propose new deep neural network architectures that lay their foundation on established models such as SRCNN, SESRCNN, EDSR, SRGAN. These models are well-known for the task of Super Resolution (SR), a technique proven in recent years to be effective in generating high-resolution images from low-resolution counterparts. The proposed models perform 10x upscaling of the low-resolution Sentinel-3 images, using Landsat 8 data as the target, leveraging the strength of Sentinel-3's temporal resolution and Landsat 8's spatial resolution. The performance of the models is evaluated based on metrics including PSNR, MAE, and SSIM and a comparison between the results is provided.