Towards High-Resolution Alignment and Super-Resolution of Multi-Sensor Satellite Imagery

Philip Wootaek Shin, Vishal Gaur, Rahul Ramachandran, Manil Maskey, Jack Sampson, Vijaykrishnan Narayanan, Sujit Roy · 2025

Read Towards High-Resolution Alignment and Super-Resolution of Multi-Sensor Satellite Imagery by Philip Wootaek Shin, Vishal Gaur, Rahul Ramachandran, Manil Maskey, Jack Sampson, Vijaykrishnan Narayanan, Sujit Roy (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-23150.txt.

Introduction

Two satellite fleets watch Earth at split levels of detail, and work that joins them must first line their pictures up and sharpen the coarse ones. Philip Shin and six co-authors build an early frame for that join in 2025. They work with a matched data set that pairs views at thirty metres per pixel with views at ten.

Past sharpening tools train on fake coarse views made by blurring fine ones, which does not match real sensor gaps. The new frame aligns the real pair and lifts the coarse view toward the fine one. Tests by score and by eye back the method. Sharper, well-aligned views feed maps of crops, floods, and land change. The work is a first step, framed by its authors as a base to build on.

Catalog record

CREATORPhilip Wootaek Shin, Vishal Gaur, Rahul Ramachandran, Manil Maskey, Jack Sampson, Vijaykrishnan Narayanan, Sujit Roy
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-23150

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Towards High-Resolution Alignment and Super-Resolution of Multi-Sensor Satellite Imagery Philip Wootaek Shin, Vishal Gaur, Rahul Ramachandran, Manil Maskey, Jack Sampson, Vijaykrishnan Narayanan, Sujit Roy Abstract High-resolution satellite imagery is essential for geospatial analysis, yet differences in spatial resolution across satellite sensors present challenges for data fusion and downstream applications. Super-resolution techniques can help bridge this gap, but existing methods rely on artificially downscaled images rather than real sensor data and are not well suited for heterogeneous satellite sensors with differing spectral, temporal characteristics. In this work, we develop a preliminary framework to align and upscale Harmonized Landsat Sentinel 30m(HLS 30) imagery using Harmonized Landsat Sentinel 10m(HLS10) as a reference from the HLS dataset. Our approach aims to bridge the resolution gap between these sensors and improve the quality of super-resolved Landsat imagery. Quantitative and qualitative evaluations demonstrate the effectiveness of our method, showing its potential for enhancing satellite-based sensing applications. This study provides insights into the feasibility of heterogeneous satellite image super-resolution and highlights key considerations for future advancements in the field. 1 Introduction High-resolution remote sensing imagery is essential for various geospatial applications. However, publicly available satellite datasets often provide images at different resolutions and scales, even for the same geographical region. While deep learning-based super-resolution techniques have demonstrated potential in enhancing spatial quality, most existing approaches rely on synthetic low-resolution images generated through conventional downscaling methods such as Lanczos [ 7 ] or bicubic interpolation. These models often overlook spatial and temporal variations, including differences in cloud coverage, and assume perfect alignment between high-resolution and low-resolution image pairs, which may not reflect real-world remote sensing conditions. Furthermore, datasets containing heterogeneous satellite images from different sensors that are precisely aligned are limited, with the WorldStrat dataset [ 5 ] being one of the few available resources. There has been little research on super-resolving images from distinct satellite sensors with differing native resolutions by leveraging one sensor as a reference to upscale another. In this work, we explore the challenges and opportunities associated with using heterogeneous RGB sensing data for super-resolution. Contributions we make are three-fold: • We develop a preliminary system for aligning and super-resolving satellite images from heterogeneous sensors, specifically using HLS10 as a reference to upscale HLS30. Unlike traditional super-resolution approaches that rely on artificially downscaled images, our method leverages real heterogenous sensor data with distinct resolutions and characteristics. • We incorporate Histogram Matching (HM) [ 8 ] and Feature Distribution Matching (FDM) [ 1 ] as preprocessing steps to mitigate spectral discrepancies between heterogeneous sensors before applying super-resolution refinement. This improves the performance of diffusion-based refinement models, such as SR3, which otherwise struggle with cross-sensor variations. • We curate a custom dataset of co-registered multi-spectral imagery from HLS 30 and HLS 10 across 18 diverse terrain regions with minimal cloud coverage. This dataset facilitates multi-channel super-resolution beyond RGB, incorporating SWIR1, SWIR2, and NIR channels, and enables exploration of training on surface reflectance values which are calculated from the raw digital numbers captured by the satellite sensors, for advancing multi-spectral super-resolution techniques. 2 Related Work We provide a brief overview of three pertinent research areas. First, we discuss satellite image upscaling, highlighting the approaches that inspired our methodology. Next, we introduce satellite datasets, with a focus on the datasets used for testing in our study. Finally, we describe our image alignment strategy, which addresses the challenges of processing heterogeneous sensor data from different satellite sources. 2.1 Remote Sensing Upscaling Remote Sensing Super-Resolution (RSSR) focuses on high-resolution (HR) reconstruction from one or more low-resolution (LR) images, facilitating critical tasks such as object detection and semantic segmentation in satellite imagery [ 20 , 12 , 22 , 2 ] . Additionally, pan-sharpening techniques, widely used in remote sensing, aim to fuse panchromatic images (high spatial resolution but low spectral resolution) with multispectral images (lower spatial resolution but higher spectral resolution) to produce HR multispectral images, thus enhancing spatial and spectral details [ 23 , 14 , 9 ] . After empirical trials, we selected the TESR [ 2 ] as our backbone architecture. TESR’s combination of Vision Transformer-based and diffusion-based refinement stages makes it particularly well-suited for addressing the challenges posed by heterogeneous data. 2.2 Dataset Remote sensing datasets play a crucial role in advancing research and applications in satellite imagery analysis. Among the most widely used datasets are Landsat and Sentinel, which provide multispectral imagery for a variety of Earth observation tasks [ 17 ] . The WorldStrat [ 5 ] dataset represents a significant step forward in providing aligned imagery from different sensors. Unlike Landsat and Sentinel, which focus on continuous data acquisition from a single sensor type, WorldStrat includes images from heterogeneous sensors (e.g., SPOT-6/7 and Sentinel-2) covering the same geographical regions. Harmonized Landsat‑8 and Sentinel‑2 (HLS) [ 17 ] surface reflectance dataset is a NASA initiative to create a virtual constellation combining Landsat‑8 and Sentinel‑2 products, harmonized to surface reflectance with consistent atmospheric correction, BRDF normalization, spatial co‑registration, and band‑matching [ 13 ] , [ 4 ] 2.3 Image alignment Conventional image alignment methods, such as Histogram Matching (HM) [ 8 ] , are widely used in image processing to establish a monotonic mapping between the histograms of two images, ensuring their intensity distributions are aligned. Additionally, deep learning techniques [ 10 , 24 ] have been developed to address more complex alignment tasks, such as correcting overexposed or underexposed images by learning non-linear transformations. Another prominent approach is statistical matching [ 1 ] , which uses statistical metrics to align images by matching their distributions or features, making it particularly effective for scenarios involving complex variations between datasets. 2.4 Surface reflectance To convert Digital Number (DN) to surface reflectance, we first compute the spectral radiance: L_{\lambda}=G\cdot\text{DN}+B (1) where L_{\lambda} is the spectral radiance (in W m -2 sr -1 \mu m -1 ), G is the radiometric gain, B is the bias (offset), and DN is the digital number from the sensor. Top-of-atmosphere (TOA) reflectance is then computed as: \rho_{\text{TOA}}=\frac{\pi\cdot L_{\lambda}\cdot d^{2}}{E_{\text{sun},\lambda}\cdot\cos(\theta_{s})} (2) where \rho_{\text{TOA}} is the TOA reflectance (unitless), d is the Earth–Sun distance in astronomical units, E_{\text{sun},\lambda} is the mean solar exoatmospheric irradiance, and \theta_{s} is the solar zenith angle. An optional atmospheric correction yields surface reflectance: \rho_{\text{surf}}=\frac{\rho_{\text{TOA}}-\rho_{\text{path}}}{T_{s}\cdot T_{v}} (3) where \rho_{\text{surf}} is the surface reflectance, \rho_{\text{path}} is the atmospheric path reflectance, and T_{s} and T_{v} are the atmospheric transmittance from the sun to the surface and from the surface to the sensor, respectively. Beyond multi-sensor fusion, surface reflectance values are foundational in numerous geoscience applications. For example, they are widely used in vegetation monitoring through indices such as Normalized Difference Vegetation Index(NDVI) and Enhanced Vegetation Index(EVI) [ 6 ] , in land cover and land use classification [ 25 ] , and in surface water and flood mapping using indices like Normalized Difference Water Index(NDWI) and Modified Normalized Difference Water Index(MNDWI) [ 21 ] . Additionally, surface reflectance is vital in soil and mineral mapping, where the spectral reflectance curves help distinguish between different surface materials [ 3 ] . In climate studies, accurate surface reflectance is used to estimate surface albedo, which influences energy balance models and carbon flux estimations [ 16 ] . Because surface reflectance captures intrinsic surface

Read the complete text