Impact of Super-Resolution on Semantic Segmentation in UAV Remote Sensing

Semantic segmentation (SS) of aerial images is an essential task in many Earth observation applications. Yet, the influence of image resolution on segmentation performance is still not well explored, especially in scenarios where only low-resolution data are available due to operational or cost constraints. In this work, we present a systematic empirical evaluation of super-resolution (SR) as an independent preprocessing stage for SS of unmanned aerial vehicle (UAV) imagery. We evaluate the impact of diverse SR proposals, based on convolutional neural networks (CNNs), transformers, and hybrid architectures across multiple scaling factors ( ×2 , ×4 , and ×8 ). To carry out this study, we design a set of experiments on the UAVid and VDD benchmark datasets using synthetically degraded LR inputs generated through Gaussian smoothing and downsampling, analyzing the effect of image resolution on segmentation, the potential of SR to enhance performance, and the relationship between SR reconstruction fidelity and segmentation outcomes. These experiments reveal how image resolution affects segmentation and how SR influences performance, providing guidance for applying SR in Earth observation tasks.