Can Location Embeddings Enhance Super-Resolution of Satellite Imagery?

TL;DR

This work addresses the challenge of enhancing the spatial resolution of publicly available satellite imagery by incorporating geographic context through location embeddings. It proposes a location-guided SR framework built on an ESRGAN backbone with SatCLIP embeddings, cross-attention conditioning, and a location-matching discriminator to enforce geographic consistency, along with local padding to mitigate tiling artifacts. The approach achieves strong CLIP-aligned perceptual quality and improves downstream building-segmentation performance compared to baselines, though traditional SR metrics do not always reflect downstream utility, and geographic generalization remains a challenge. Overall, the method offers a practical path toward more geographically faithful, high-resolution satellite imagery suitable for urban planning and disaster response tasks.

Abstract

Publicly available satellite imagery, such as Sentinel- 2, often lacks the spatial resolution required for accurate analysis of remote sensing tasks including urban planning and disaster response. Current super-resolution techniques are typically trained on limited datasets, leading to poor generalization across diverse geographic regions. In this work, we propose a novel super-resolution framework that enhances generalization by incorporating geographic context through location embeddings. Our framework employs Generative Adversarial Networks (GANs) and incorporates techniques from diffusion models to enhance image quality. Furthermore, we address tiling artifacts by integrating information from neighboring images, enabling the generation of seamless, high-resolution outputs. We demonstrate the effectiveness of our method on the building segmentation task, showing significant improvements over state-of-the-art methods and highlighting its potential for real-world applications.

Figures (7)

• Figure 1: Illustration of the super-resolution pipeline and its application on an area of the city of Tashkent, Uzbekistan. (a) Sentinel-2 input image with low resolution, (b) super-resolved output, and (c) building segmentation result.
• Figure 2: Visual comparison of satellite images from geographically diverse regions using the proposed super-resolution pipeline. The left image in each pair represents the low-resolution Sentinel-2 input, while the right shows the corresponding super-resolved output, demonstrating the generalization of our method across different global urban landscapes.
• Figure 3: Summary of the proposed architecture. The input image from Sentinel-2 is processed through a shallow feature extraction network (CNN), followed by the RRDB block for deep feature extraction. Location embeddings, extracted using the SatCLIP module, are projected via MLP layers (purple) and injected into the network during the upsampling phase. The self-attention (yellow) and cross-attention (blue) mechanisms facilitate the integration of the image features and location embeddings, enabling the network to produce a super-resolved image that is both high in visual quality and geographically consistent.
• Figure 4: Overview of the location-matching Discriminator. The input consists of an image paired with real and false geographic coordinates, processed through the SatCLIP module. The image and the location embeddings are then passed through the discriminator network. The output of the discriminator is a binary classification (real or fake) based on whether the location matches the input image. The SatCLIP location embeddings guide the discriminator, enforcing geographic consistency in the image generation process.
• Figure 5: Visualization of the data split by UTM zones. The train set includes samples from UTM zones 10-17, while the validation and test set consists of samples from UTM zones 13N, 18N, and 19N. These zones were selected to evaluate how well the model generalizes across different geographic regions.