Abstract. Low-resolution imagery is a major constraint for remote sensing tasks (e.g., urban land cover detection) where accurate classification of buildings, roads, vegetation, and small objects is required. Deep learning-based segmentation models are highly sensitive to image quality, resulting in degraded performance on low-resolution inputs. Super-resolution (SR) techniques offer a promising solution by enhancing image fidelity to support downstream tasks. This work applied MAPSRNet, a Multi-Attention Pyramid SR Network to aerial images used for multi-class land cover detection. Evaluated on the ISPRS Potsdam dataset, MAPSRNet achieves state-of-the-art SR performance with PSNR of 32.92 dB and SSIM of 0.87, outperforming existing methods such as SRCNN (31.54 dB, 0.83) and DRRN (31.03 dB, 0.82) while maintaining competitive inference speed. Beyond image quality, MAPSRNet significantly improves multi-class land cover segmentation when integrated with a ConvNeXtV2-based U-Net, achieving an overall accuracy of 80.60%, mean IoU of 62.54%, and FwIoU of 68.34%, surpassing not only low-resolution inputs (Overall Accuracy: 65.28%, mIoU: 40.20%, FwIoU: 50.12%) but also high-resolution(HR) ones (Overall Accuracy: 80.50%, mIoU: 62.40%, FwIoU: 68.01%), especially in certain classes such as impervious surface and clutter. These results demonstrate that perceptual and structural fidelity, rather than pixel-level similarity, can drive superior performance in urban land cover segmentation. MAPSRNet offers a practical solution for scenarios where HR imagery is limited or unavailable, highlighting its potential for large-scale remote sensing applications.
Satellite imagery often suffers from limited spatial resolution and, in many cases, high acquisition costs. These factors restrict their use in applications such as urban monitoring, land management, and wildlife studies. This work proposes an AI-based super-resolution approach that leverages high resolution aerial imagery to train a Generative Adversarial Network. Specifically, the ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) architecture is adapted and trained using aerial orthophotos, enabling the transfer of learned spatial representations to low-resolution satellite images. The trained model is evaluated on satellite image patches at 2 and 4 super-resolution scales. Performance is assessed using structural, perceptual, and chromatic metrics, including SSIMY, MS-SSIM, LPIPS and CIEDE2000. The results show clear improvements, with increased sharpness, enhanced edge definition, and consistent reconstruction of urban structures and terrain features. From a quantitative perspective, the 2 scale achieves the best overall metric values, while the 4 scale maintains stable and meaningful performance despite the higher reconstruction difficulty. These findings demonstrate the feasibility of transferring super-resolution capabilities from aerial images to satellite imagery, even in the presence of spectral and geometric differences between acquisition domains. Overall, this study provides a solid foundation for the development of low-cost, AI-driven satellite image super-resolution models and outlines future research directions focused on dataset expansion, domain adaptation strategies, and sensor-specific architectural improvements.
Abstract. High-resolution Earth observation data are crucial for applications such as agriculture, urban planning, and environmental monitoring. Although commercial satellites provide sub-meter imagery, open-access alternatives like Sentinel-2 are limited to resolutions around 10 m ground sampling distance, which is insufficient for many tasks. In this work, we investigate image super-resolution as a method to bridge this gap, enhancing downstream performance on freely available satellite data. We leverage two 16-bit single-band datasets, consisting of Sentinel-2 (20 m→10 m) and VENμS (10 m→5 m) images, to train and benchmark state-of-the-art SR methods, including transformer- and diffusion-based approaches, across multiple dataset mixes. These models are evaluated quantitatively using reference-based metrics (PSNR, SSIM) using ground-truth and no-reference scores (FID, NIQE) for native upscaling from 20 m→10 m and 10 m→5 m. We observe that different SR architectures present trade-offs between standard quantitative metrics and perceptual image quality. We further assess their impact on a practical downstream task: field boundary detection from Sentinel-2 imagery. Our experiments demonstrate that SR pre-processing improves quantitative fidelity and downstream task performance, enabling low-resolution satellites to compete more effectively with commercial imagery.
High-resolution aerial imagery plays a significant role in many fields such as urban planning, environmental monitoring, weather prediction, disaster management, change detection and map generation. However, acquiring high-resolution data is often limited by sensor capacities and cost constraints. Moreover, in aerial remote sensing platforms, raw images are compressed during downlink transmission to reduce bandwidth requirements, energy consumption and storage capacity. Lossy image compression algorithms including JPEG degrade image quality, causing artifacts such as blurring, blocking and ringing. In this work, we propose a two-stage framework that employs FBCNN for artifact removal and ESRGAN for super-resolution reconstruction in aerial imagery. We construct our test set based on the SODA-A dataset. Experimental results show that FBCNN achieves high PSNR, SSIM and PSNR-B values across different JPEG quality factors, with PSNR ranging from 28.63 dB to 35.09 dB, SSIM from 0.769 to 0.930, and PSNR-B from 28.40 dB to 34.23 dB. Building upon the outputs of FBCNN, ESRGAN further enhances perceptual quality while maintaining strong quantitative performance, achieving PSNR values from 25.30 dB to 27.99 dB and SSIM values from 0.610 to 0.696 across different quality factors.
The detection of small objects in aerial traffic scenes is vital for advancing intelligent transportation systems (ITS), yet traditional methods often suffer from scale variation, background complexity, and insufficient context utilization. To address these challenges, we propose SOD-MSC, a YOLOv8-based small object detector designed for remote sensing images. SOD-MSC integrates a global multi-scale (GMS) module to enhance global perception and multi-scale fusion, a lightweight fusion feature pyramid network (LFFPN) to strengthen spatial correlations and suppress background noise, and an adaptive RT-DETR-based detection head with Inner_SIoU loss to improve localization accuracy and convergence. Experiments on VisDrone2019-DET and HazyDet demonstrate that SOD-MSC achieves 39.9% and 49% on mAP@0.5, with only 5.22 M parameters and 11.2 GFLOPs, surpassing several baselines and state-of-the-art models in balancing accuracy and efficiency. These results highlight the effectiveness of SOD-MSC for dense small object detection in complex traffic environments, with promising applications in congestion monitoring, vehicle counting, and related ITS tasks.
Road segmentation from satellite imagery is critical for urban planning and transportation analysis, but is often limited by the low spatial resolution of publicly available data and the high cost of high-resolution alternatives. This study evaluates the impact of super-resolution (SR) on urban road network extraction in Dubai. Sentinel-2 (S2) imagery with a native resolution of 10m was enhanced to approximately 1m using the Sentinel-2 Deep Resolution 3 (S2DR3) model, and road segmentation was performed using the SAM-LoRA model. The model was trained on 105 km2 of Dubai city using a small set of manually annotated road segments and evaluated on an independent region. Despite limited training annotations, the model achieved robust performance, demonstrating the effectiveness of combining SR preprocessing with the SAM-LoRA architecture. Due to the lack of high-resolution reference imagery, no-reference image quality metrics were used to assess SR output. Quantitative results indicate that SR improves segmentation performance, achieving an overall accuracy of 0.91, Kappa of 0.81, IoU of 0.82, and F1-score of 0.90, compared to 0.68, 0.35, 0.36, and 0.54, respectively, for the original S2 imagery. Qualitative analysis shows that SR generates smooth, continuous road networks, whereas outputs from the original S2 imagery were fragmented and imprecise. Beyond the testing region, the SR imagery and trained model were applied to the entire city of Dubai (∼ 1700 km2), producing a city-wide road network and demonstrating strong generalization. These findings confirm that SR preprocessing can improve the accuracy of extracting urban road networks from low-resolution satellite images.