A benchmark dataset for Landsat-to-Sentinel image generation using deep learning-driven super-resolution techniques

IF 2.8 3区地球科学 Q2 ASTRONOMY & ASTROPHYSICS

Advances in Space Research Pub Date : 2026-03-15 Epub Date: 2026-01-21 DOI:10.1016/j.asr.2026.01.049

Peijuan Wang , Samet Aksoy , Elif Sertel

{"title":"A benchmark dataset for Landsat-to-Sentinel image generation using deep learning-driven super-resolution techniques","authors":"Peijuan Wang , Samet Aksoy , Elif Sertel","doi":"10.1016/j.asr.2026.01.049","DOIUrl":null,"url":null,"abstract":"<div><div>High-resolution satellite imagery plays a vital role in accurately analyzing surface changes, vegetation dynamics, and land cover transitions for environmental monitoring and Earth science applications. While the Landsat satellite series provides long-term, high-coverage time-series data—essential for studying large-scale phenomena such as deforestation, urban expansion, and agricultural transformation—its 30-meter spatial resolution often falls short in applications requiring finer detail. To address this limitation, this study introduces Land2Sent, a novel remote sensing super-resolution dataset specifically designed for the Landsat 8/9 to Sentinel-2A/B image enhancement task. The Land2Sent dataset aims to upscale Landsat imagery from 30 m to 10 m by utilizing the higher-resolution Sentinel-2 images as reference. Both normalized 4-band (R, G, B, NIR) images and original 16-bit 4-band images are included to assess the impact of bit depth on model performance. Using this dataset, ten state-of-the-art deep learning models are evaluated for their ability to reconstruct super-resolved images from low-resolution Landsat inputs. The performance of these models is assessed using quantitative metrics across the full dataset, as well as through visual inspection and Normalized Difference Vegetation Index (NDVI) analysis of selected image patches.</div></div>","PeriodicalId":50850,"journal":{"name":"Advances in Space Research","volume":"77 6","pages":"Pages 6855-6880"},"PeriodicalIF":2.8000,"publicationDate":"2026-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Advances in Space Research","FirstCategoryId":"89","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0273117726000748","RegionNum":3,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2026/1/21 0:00:00","PubModel":"Epub","JCR":"Q2","JCRName":"ASTRONOMY & ASTROPHYSICS","Score":null,"Total":0}

引用次数: 0

Abstract

High-resolution satellite imagery plays a vital role in accurately analyzing surface changes, vegetation dynamics, and land cover transitions for environmental monitoring and Earth science applications. While the Landsat satellite series provides long-term, high-coverage time-series data—essential for studying large-scale phenomena such as deforestation, urban expansion, and agricultural transformation—its 30-meter spatial resolution often falls short in applications requiring finer detail. To address this limitation, this study introduces Land2Sent, a novel remote sensing super-resolution dataset specifically designed for the Landsat 8/9 to Sentinel-2A/B image enhancement task. The Land2Sent dataset aims to upscale Landsat imagery from 30 m to 10 m by utilizing the higher-resolution Sentinel-2 images as reference. Both normalized 4-band (R, G, B, NIR) images and original 16-bit 4-band images are included to assess the impact of bit depth on model performance. Using this dataset, ten state-of-the-art deep learning models are evaluated for their ability to reconstruct super-resolved images from low-resolution Landsat inputs. The performance of these models is assessed using quantitative metrics across the full dataset, as well as through visual inspection and Normalized Difference Vegetation Index (NDVI) analysis of selected image patches.

查看原文本刊更多论文

使用深度学习驱动的超分辨率技术生成Landsat-to-Sentinel图像的基准数据集

高分辨率卫星图像在准确分析地表变化、植被动态和土地覆盖变化方面发挥着至关重要的作用，为环境监测和地球科学应用提供了基础。虽然Landsat卫星系列提供了长期、高覆盖的时间序列数据，这对于研究森林砍伐、城市扩张和农业转型等大规模现象至关重要，但其30米的空间分辨率在需要更精细细节的应用中往往不足。为了解决这一限制，本研究引入了Land2Sent，这是一个专门为Landsat 8/9到Sentinel-2A/B图像增强任务设计的新型遥感超分辨率数据集。Land2Sent数据集旨在利用更高分辨率的Sentinel-2图像作为参考，将Landsat图像从30米提升到10米。归一化的4波段（R， G， B， NIR）图像和原始的16位4波段图像都被包括在内，以评估位深度对模型性能的影响。使用此数据集，评估了10个最先进的深度学习模型从低分辨率Landsat输入重建超分辨率图像的能力。这些模型的性能使用整个数据集的定量指标进行评估，以及通过视觉检查和选定图像斑块的归一化植被指数（NDVI）分析。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Advances in Space Research 地学天文-地球科学综合

CiteScore

5.20

自引率

11.50%

发文量

800

审稿时长

5.8 months

期刊介绍： The COSPAR publication Advances in Space Research (ASR) is an open journal covering all areas of space research including: space studies of the Earth''s surface, meteorology, climate, the Earth-Moon system, planets and small bodies of the solar system, upper atmospheres, ionospheres and magnetospheres of the Earth and planets including reference atmospheres, space plasmas in the solar system, astrophysics from space, materials sciences in space, fundamental physics in space, space debris, space weather, Earth observations of space phenomena, etc. NB: Please note that manuscripts related to life sciences as related to space are no more accepted for submission to Advances in Space Research. Such manuscripts should now be submitted to the new COSPAR Journal Life Sciences in Space Research (LSSR). All submissions are reviewed by two scientists in the field. COSPAR is an interdisciplinary scientific organization concerned with the progress of space research on an international scale. Operating under the rules of ICSU, COSPAR ignores political considerations and considers all questions solely from the scientific viewpoint.