Enhancing Sentinel-2 Images using Super Resolution

High-resolution satellite imagery plays a crucial role in remote sensing, enabling a wide range of applications such as land-use monitoring, urban planning, precision agriculture, and environmental change detection. Super-resolution is a computer vision task aimed at reconstructing high-resolution images from their low-resolution counterparts. Applying it to multi-temporal, multi-spectral satellite imagery, such as that provided by Sentinel-2, can yield promising results; however, it is often not suitable as input for downstream tasks, such as land-cover segmentation. Multi-task learning offers a viable alternative, providing improved regularization and better preservation of spatial structures when generating high-resolution land-cover maps. Moreover, information carried by shared feature representations proves more effective than relying on super-resolved RGB images. This work makes use of two datasets: one specifically designed for super-resolution, and another suited for land-cover segmentation. After properly adapting the FLAIR-2 dataset, it investigates super-resolution architectures with the goal of extending and employing them to perform both tasks in a single forward pass. The main objective is to obtain 4x super-resolved land-cover maps, supported by weighted combined loss functions. The experiments involve fully convolutional networks (SRCNN, RCAN), generative adversarial networks (ESRGAN), and vision transformers (SwinIR). The study also explores the performance of pre-trained models, evaluates the benefits of parameter-efficient fine-tuning techniques (such as LoRA), and examines adversarial multi-task learning strategies. Finally, quantitative and qualitative results are presented, showing that the proposed multi-task approach improves segmentation performance over single-task baselines, particularly in the preservation of spatial structures. A discussion of the advantages, limitations, and future research directions for this approach is also provided.