How Does Super-Resolution for Satellite Imagery Affect Different Types of Land Cover? Sentinel-2 Case
Anna Malczewska, Maciej Wielgosz
Abstract
In the dynamic field of satellite imagery, the significance of super-resolution techniques, grounded on advanced deep learning methods, is paramount. A thorough understanding and remediation of the distinct challenges posed by various land cover types for image resolution enhancement form the essence of this research. This work diligently employs two unique neural networks, SRCNN and SwinIR Transformer, to scrutinize their varying impacts on a range of land cover types, ensuring a detailed and comprehensive exploration. This study transcends the mere enhancement of the Sentinel-2 dataset's resolution from 20m/pix to 10m/pix. It ambitiously seeks to excavate the intricate trends inherent to different land cover types and their corresponding interactions with super-resolution processes. The application of neural networks on 255x254 pixel patches, covering six dominant types — forests, large fields, small fields, urban, suburban, and mixed — highlights substantial variations in metrics, underlining the individual interactions of each land cover type with super-resolution techniques. A comprehensive accuracy assessment is meticulously conducted, employing an array of metrics and frequency domains to shed light on the nuanced differences and provide vital insights for optimizing each land cover type's super-resolution approaches. Notably, the PSNR metric reveals significant disparities, particularly in the ‘forest’ and ‘urban’ categories for both SRCNN and SwinIR. According to the PSNR metric, the ‘forest’ class yielded the best results with 66.06 for SRCNN and 67.00 for SwinIR, while the ‘urban’ class marked the lowest with 55.09 and 57.02 respectively, reinforcing the critical nature of this study.