Cloud-Free Sentinel-2 Mosaic Generation for the Republic of Korea
Donggyu Kim1, Seungchan Lee2, Chuluong Choi3

1 Bachelors student, Division of Earth and Environmental System Science, Pukyong National University, Republic of Korea
2 Masters student, Division of Earth and Environmental System Science (Spatial Information Engineering), Pukyong National University, Republic of Korea
3 Professor, Division of Earth and Environmental System Science (Spatial Information Engineering), Pukyong National University, Republic of Korea
*cuchoi[at]pknu.ac.kr

Abstract

In large-scale satellite imagery mosaicking, the presence of clouds obstructs analysis and necessitates cloud removal- however, uniform cloud elimination often generates missing data regions that can bias downstream applications. This study develops an automated pipeline leveraging GPU-accelerated processing to produce cloud-free mosaic images across the Republic of Korea using Sentinel-2 satellite imagery. A total of 1,352 Level-1C Sentinel-2 scenes collected from March through July 2025 covered 22 ESA-standard tiles of the Korean Peninsula. First, datastrip-derived scene fragments were spatially coregistered and merged into unified GeoTIFFs. Path-dependent variations in spatial resolution and coordinate reference systems were reconciled via nearest-neighbor resampling to ensure consistent pixel alignment. In the primary cloud-screening stage, initial binary masks were generated using cloud probability thresholds based on pixel-level cloud scores- unused spectral bands were removed and cloud pixels were masked out. For secondary cloud separation, monthly Normalized Difference Vegetation Index (NDVI) was computed for each scene to identify the clearest image as the master, with the remaining scenes designated as slaves. Slave scenes informed NDVI-based masking and only scenes achieving a minimum 45 percent no-cloud land ratio qualified for master selection. To fill missing data in the master image, pixel-wise linear regression coefficients precomputed from slave scenes were applied via a linear regression formula to estimate absent reflectance values, followed by sequential spatial interpolation. Residual gaps were then filled using a neighbor-based gap-filling algorithm that honors local spatial continuity. Finally, for months where all Sentinel-2 scenes were cloud-obscured, auxiliary imagery from alternative satellite sources was tiled and resampled into the workflow. The proposed method enhances data completeness and processing efficiency, offering a robust solut

Keywords: mosaic- cloud elimination- fill missing data- spatial interpolation- gap-filling

Topic: Topic A: General Remote Sensing