Earth observation satellite technology has enabled the acquisition of optical satellite imagery with various spatiotemporal resolutions. While sub-meter resolution satellite imagery enables precise land monitoring, its narrow swath width and data volume constraints limit the construction of time-series data. To address this trade-off between spatial and temporal resolution, spatiotemporal fusion methods have been developed. However, existing research has primarily focused on medium-to-low resolution satellites like Landsat-MODIS, with high-resolution applications limited to 2-3m resolution. Furthermore, most studies have only considered seasonal vegetation changes, leaving performance unverified in complex urban areas.

This study established and analyzed four research questions regarding potential issues in applying spatiotemporal fusion to high-resolution imagery: 1) whether existing methods can capture fine-scale changes in high-resolution imagery, 2) whether high-resolution information is preserved during fusion, 3) how viewing angle affects the results, and 4) whether building classification is feasible using fused products. We applied STARFM, EDCSTFN, and GAN-STFM methods to 0.5m CAS500-1 and daily PlanetScope imagery.

Analysis revealed that STARFM performed well in unchanged areas but failed to reflect changes and exhibited blurring. Learning-based methods better captured changes but showed significant quality degradation when objects were created, while performing better when objects were removed due to high-frequency information loss. Regarding high-resolution information preservation, results varied by object size and characteristics. Learning-based methods lost high-frequency information for small objects due to convolution operations but clearly reconstructed features smaller than the kernel size, like road markings.

Viewing angle effects influenced shadow regions and building tilt processing. EDCSTFN showed high sensitivity to off-nadir angles in training data, while GAN-STFM showed lower dependence. Learning-based methods could generate missing information in shadow regions and reduced building tilt effects. Building classification performance varied with building size and arrangement. While performance matched the original imagery for large, well-spaced buildings, it degraded in dense urban areas and changed regions. Convolution operations in learning-based methods led to boundary ambiguity and overestimation, while STARFM's blurring caused frequent misclassification at building-road boundaries, indicating that spectral distortion can reduce classification accuracy despite preserved visual boundaries.

This study contributes by identifying potential issues, limitations, and their causes when applying spatiotemporal fusion to high-resolution satellite imagery. Future research should validate across diverse sites and resolution differences, and develop new fusion methods considering high-resolution imagery characteristics.



Keywords : 시공간 융합, 고해상도 위성영상, 초소형군집위성, 도시 모니터링, 건물 분류