Recent advances in single-and multi-image superresolution have revealed the limitations of classical image similarity metrics (like peak signal-to-noise ratio), as they often fail to align with human perception when evaluating the visual quality of super-resolved outputs.In this paper, we explore how to exploit keypoint-based metrics to evaluate super-resolution image quality.Specifically, we explore two correlated metrics: (i) a multiscale index proposal measure capturing salience of keypoints, and (ii) a repeatability metric quantifying how consistently the corresponding keypoints are identified in super-resolved and ground-truth images.Experiments on several simulated and real-world datasets show that the repeatability correlates with subjective judgments, and multi-scale index proposal can be helpful for difficult datasets when other metrics are insufficient.

Insufficient spatial resolution remains a serious obstacle for many potential applications of Sentinel-2 images. Therefore, a number of super-resolution techniques have been proposed, including single-image and multi-image approaches. While the former may hallucinate (rather than reconstruct) the image details, the latter are underpinned with information fusion that helps address these issues. However, given the five-days-long revisit time, a multi-temporal series of Sentinel-2 images contains temporal changes that affect the performance of multi-image super-resolution. In this paper, we propose a solution that allows for selecting a single image in the series to specify the point in time, at which the scene should be reconstructed. We report the results of our extensive experiments which confirm that our approach leads to improving the reconstruction quality, while preserving the temporal consistency. Importantly, the proposed solutions are potentially applicable to a variety of existing super-resolution techniques and they are not limited to Sentinel-2 images.

Multispectral Sentinel-2 images are a valuable source of Earth observation data, however spatial resolution of their spectral bands limited to 10 m, 20 m, and 60 m ground sampling distance remains insufficient in many cases. This problem can be addressed with super-resolution, aimed at reconstructing a high-resolution image from a low-resolution observation. For Sentinel-2, spectral information fusion allows for enhancing the 20 m and 60 m bands to the 10 m resolution. Also, there were attempts to combine multitemporal stacks of individual Sentinel-2 bands, however these two approaches have not been combined so far. In this paper, we introduce DeepSent—a new deep network for super-resolving multitemporal series of multispectral Sentinel-2 images. It is underpinned with information fusion performed simultaneously in the spectral and temporal dimensions to generate an enlarged multispectral image. In our extensive experimental study, we demonstrate that our solution outperforms other state-of-the-art techniques that realize either multitemporal or multispectral data fusion. Furthermore, we show that the advantage of DeepSent results from how these two fusion types are combined in a single architecture, which is superior to performing such fusion in a sequential manner. Importantly, we have applied our method to super-resolve real-world Sentinel-2 images, enhancing the spatial resolution of all the spectral bands to 3.3 m nominal ground sampling distance, and we compare the outcome with very high-resolution WorldView-2 images. We will publish our implementation upon paper acceptance, and we expect it will increase the possibilities of exploiting super-resolved Sentinel-2 images in real-life applications.

Multi-image super-resolution is a branch of super-resolution reconstruction techniques aiming to resolve a set of lowresolution images into a high-resolution one. Unlike singleimage super-resolution, its goal is to fuse information embedded in different images depicting the same scene, usually captured at different times. Such data combination contains more high-resolution information than a single image, thus allowing for more accurate reconstruction results. One of the most critical challenges is to prepare training data for multi-image super-resolution since only a few datasets are available here, especially for satellite imaging applications. For this reason, many studies are conducted using simulated low-resolution images, but the results obtained for real-life data are often unsatisfactory. To overcome this problem, we propose a new semi-simulated approach of creating lowresolution images for training that resemble real-life ones much more accurately. We also investigate the performance of selected deep learning models trained with simulated and semi-simulated datasets and we show that the latter achieve better results when applied to real-world images.

The recent advances in Earth observation and artificial in-telligence allow us to improve the agricultural management practices through effectively exploiting the spectral, spatial, and temporal characteristics of the area of interest captured by satellite images. In this paper, we tackle the problem of extracting high-resolution (2.5-meter) cultivated land maps from Sentinel-2 multispectral images, and propose a machine learning algorithm for this task. It aggregates the spectral, spatial, and temporal features of the upsampled images, and is independent from the number of observations captured for a given scene. The experimental results, performed within the framework of the Enhanced Sentinel-2 Agriculture chal-lenge show that our technique manifests high generalization abilities over the unseen data and elaborates high-quality cul-tivated land maps. Finally, utilizing this algorithm led us to taking the $6^{\text{th}}$ place in the aforementioned challenge.

Super-resolution consists in reconstructing a high-resolution image from single or multiple low-resolution observations. Deep learning has been reported extremely successful for single-image super-resolution, but its applications to the multiple-image scenarios are limited due to the challenges that arise from feeding a network with a stack of images with sub-pixel translations. In this paper, we introduce Magnet—a new graph neural network that benefits from representing the input low-resolution images as a graph. This enables us to exploit the sub-pixel shifts among the input images while preserving the original low-resolution pixel values for feature extraction and information fusion. Despite a relatively simple architecture, Magnet outperforms the state-of-the-art methods for multiple-image super-resolution, and due to the flexible graph representation, it allows for using a variable number of low-resolution images for reconstruction.

Super-resolution (SR) reconstruction is a common term for techniques aimed at generating a high-resolution image from a single low-resolution image or multiple images showing the same scene. Multiple-image SR benefits from data fusion which allows for more accurate reconstruction of the underlying high-resolution information. Deep learning is extensively used for single-image SR, but its application to multiple-image SR is much less explored. Recently, several deep networks were proposed to enhance Proba-V images, and in this paper, we focus on employing them to super-resolve the Sentinel-2 images. In particular, we investigate the influence of the training data, including real and simulated low-resolution images, on the final SR outcome. Also, we make the simulated data publicly available.

The use of deep-learned features has recently allowed for improving the performance of skin detection and segmentation. In particular, fully-convolutional U-Nets, proposed for segmenting medical images, occurred to be extremely effective here. However, the spatial context, which is rather narrow for U-Nets, may be more important for skin segmentation than for segmenting other image structures. We propose Skinny-a lightweight U-Net-based architecture that extends the range of multi-scale analysis. The results of our experiments indicate that Skinny outperforms the state-of-the-art skin segmentation techniques, rendering the F-score of 92.3% and 94.9% for the ECU and HGR datasets, respectively.