The detection of building footprints and road networks has many useful applications including the monitoring of urban development, real-time navigation, etc. Taking into account that a great deal of human attention is required by these remote sensing tasks, a lot of effort has been made to automate them. However, the vast majority of the approaches rely on very high-resolution satellite imagery (<2.5 m) whose costs are not yet affordable for maintaining up-to-date maps. Working with the limited spatial resolution provided by high-resolution satellite imagery such as Sentinel-1 and Sentinel-2 (10 m) makes it hard to detect buildings and roads, since these labels may coexist within the same pixel. This paper focuses on this problem and presents a novel methodology capable of detecting building and roads with sub-pixel width by increasing the resolution of the output masks. This methodology consists of fusing Sentinel-1 and Sentinel-2 data (at 10 m) together with OpenStreetMap to train deep learning models for building and road detection at 2.5 m. This becomes possible thanks to the usage of OpenStreetMap vector data, which can be rasterized to any desired resolution. Accordingly, a few simple yet effective modifications of the U-Net architecture are proposed to not only semantically segment the input image, but also to learn how to enhance the resolution of the output masks. As a result, generated mappings quadruplicate the input spatial resolution, closing the gap between satellite and aerial imagery for building and road detection. To properly evaluate the generalization capabilities of the proposed methodology, a data-set composed of 44 cities across the Spanish territory have been considered and divided into training and testing cities. Both quantitative and qualitative results show that high-resolution satellite imagery can be used for sub-pixel width building and road detection following the proper methodology.

Earth observation data is becoming more accessible and affordable thanks to the Copernicus programme and its Sentinel missions. Every location worldwide can be freely monitored approximately every 5 days using the multi-spectral images provided by Sentinel-2. The spatial resolution of these images for RGBN (RGB + Near-infrared) bands is 10 m, which is more than enough for many tasks but falls short for many others. For this reason, if their spatial resolution could be enhanced without additional costs, any posterior analyses based on these images would be benefited. Previous works have mainly focused on increasing the resolution of lower resolution bands of Sentinel-2 (20 m and 60 m) to 10 m resolution. In these cases, super-resolution is supported by bands captured at finer resolutions (RGBN at 10 m). On the contrary, this paper focuses on the problem of increasing the spatial resolution of 10 m bands to either 5 m or 2.5 m resolutions, without having additional information available. This problem is known as single-image super-resolution. For standard images, deep learning techniques have become the de facto standard to learn the mapping from lower to higher resolution images due to their learning capacity. However, super-resolution models learned for standard images do not work well with satellite images and hence, a specific model for this problem needs to be learned. The main challenge that this paper aims to solve is how to train a super-resolution model for Sentinel-2 images when no ground truth exists (Sentinel-2 images at 5 m or 2.5 m). Our proposal consists of using a reference satellite with a high similarity in terms of spectral bands with respect to Sentinel-2, but with higher spatial resolution, to create image pairs at both the source and target resolutions. This way, we can train a state-of-the-art Convolutional Neural Network to recover details not present in the original RGBN bands. An exhaustive experimental study is carried out to validate our proposal, including a comparison with the most extended strategy for super-resolving Sentinel-2, which consists in learning a model to super-resolve from an under-sampled version at either 40 m or 20 m to the original 10 m resolution and then, applying this model to super-resolve from 10 m to 5 m or 2.5 m. Finally, we will also show that the spectral radiometry of the native bands is maintained when super-resolving images, in such a way that they can be used for any subsequent processing as if they were images acquired by Sentinel-2.

Obtaining Sentinel-2 imagery of higher spatial resolution than the native bands while ensuring that output imagery preserves the original radiometry has become a key issue since the deployment of Sentinel-2 satellites. Several studies have been carried out on the upsampling of 20m and 60m Sentinel-2 bands to 10 meters resolution taking advantage of 10m bands. However, how to super-resolve 10m bands to higher resolutions is still an open problem. Recently, deep learning-based techniques has become a de facto standard for single-image super-resolution. The problem is that neural network learning for super-resolution requires image pairs at both the original resolution (10m in Sentinel-2) and the target resolution (e.g., 5m or 2.5m). Since there is no way to obtain higher resolution images for Sentinel-2, we propose to consider images from others sensors having the greatest similarity in terms of spectral bands, which will be appropriately pre-processed. These images, together with Sentinel-2 images, will form our training set. We carry out several experiments using state-of-the-art Convolutional Neural Networks for single-image super-resolution showing that this methodology is a first step toward greater spatial resolution of Sentinel-2 images.

Abstract. Copernicus program via its Sentinel missions is making earth observation more accessible and affordable for everybody. Sentinel-2 images provide multi-spectral information every 5 days for each location. However, the maximum spatial resolution of its bands is 10m for RGB and near-infrared bands. Increasing the spatial resolution of Sentinel-2 images without additional costs, would make any posterior analysis more accurate. Most approaches on super-resolution for Sentinel-2 have focused on obtaining 10m resolution images for those at lower resolutions (20m and 60m), taking advantage of the information provided by bands of finer resolutions (10m). Otherwise, our focus is on increasing the resolution of the 10m bands, that is, super-resolving 10m bands to 2.5m resolution, where no additional information is available. This problem is known as single-image super-resolution and deep learning-based approaches have become the state-of-the-art for this problem on standard images. Obviously, models learned for standard images do not translate well to satellite images. Hence, the problem is how to train a deep learning model for super-resolving Sentinel-2 images when no ground truth exist (Sentinel-2 images at 2.5m). We propose a methodology for learning Convolutional Neural Networks for Sentinel-2 image super-resolution making use of images from other sensors having a high similarity with Sentinel-2 in terms of spectral bands, but greater spatial resolution. Our proposal is tested with a state-of-the-art neural network showing that it can be useful for learning to increase the spatial resolution of RGB and near-infrared bands of Sentinel-2.