A harmonization-based framework for cross-sensor super-resolution of Sentinel-2 imagery for coastal applications

Abstract

High-resolution (HR) satellite imagery is essential for coastal mapping, ocean analysis, and long-term environmental monitoring; however, acquiring it is constrained by hardware limitations and high costs. Freely available low-resolution (LR) archives, such as those from Sentinel-2 (S2), offer global coverage and frequent revisits, but their 10 m native resolution is insufficient for fine-scale analysis. This study presents HARBOR (Harmonization-based Adaptive Resolution Boosting for Optical Remote sensing), an integrated harmonization-based workflow that combines established radiometric harmonization and sub-pixel convolutional super-resolution (SR) components and adapts them to four-band coastal S2–PlanetScope (PS) data. Rather than proposing a new network architecture, HARBOR addresses a gap in the way satellite SR is commonly assembled, where network optimization receives most of the attention and cross-sensor radiometric inconsistency is treated as a preprocessing detail. The workflow consists of four interconnected stages: (1) a same-day, close-time data preparation approach for temporally consistent cross-sensor learning, (2) statistical radiometric harmonization to reduce cross-sensor reflectance differences, (3) the synthetic generation of S 2 like imagery to expand limited supervised datasets, and (4) a customized Efficient Sub-Pixel Convolutional Neural Network (ESPCN) adapted for multispectral satellite imagery using a fused loss formulation. In a coastal case study of Galveston Island, Texas, USA, PS imagery resampled from 3 m to 5 m served as the HR reference for reconstructing S2 inputs from 10 m to 5 m. Because only 18 same-day, close-time S2 and PS pairs were available for direct cross-sensor training, a harmonization model calibrated on those pairs was used to generate synthetic S 2 like imagery from PS, expanding the supervised training set to 191 temporally consistent pairs. A component ablation separates the three elements of the workflow, namely statistical radiometric harmonization, synthetic S 2 like expansion of the training set, and the fused loss. Compared at a fixed training-set size, radiometric harmonization is the largest single contributor to reconstruction quality, ahead of the fused loss, while synthetic expansion is what makes supervised training feasible given the scarcity of real same-day pairs. The harmonization parameters and trained weights are calibrated to a single study site and require re-estimation on local same-day pairs before use elsewhere; what transfers is the end-to-end workflow itself, which integrates cross-sensor data preparation, radiometric harmonization, synthetic S 2 like generation, and neural network-based reconstruction into a single transferable pipeline, rather than the Galveston-calibrated model. Enhanced S2 archives of this kind could in future support coastal monitoring, environmental assessment, and the monitoring of Nature-Based Solutions (NbS) for coastal flood resilience, though no downstream task of that type is demonstrated.