An Improved Spatiotemporal Fusion Framework for Land-Cover Temporal Harmonization of High-Resolution Remote Sensing Images

Abstract

High-resolution remote sensing images, with fine spatial detail but limited coverage and infrequent revisits, often exhibit temporal discrepancies that hinder large-scale Earth observations. While spatiotemporal fusion (STF) methods offer a solution, they often lead to reduced spatial resolution and struggle with multisource and multitemporal image processing. To address this issue, an improved STF framework for temporal harmonization (STF-TH) was proposed. Specifically, STF-TH first applies an existing STF method for initial temporal transformation. Second, spatial resolution is recovered through spatial texture correction, referencing the fine texture of the original image. Finally, temporal color correction leverages the consistency of coarse images to further reduce temporal discrepancies among results. STF-TH was evaluated across datasets collected from different satellites, regions, and times, and validated via both qualitative and quantitative analyses at global, local, and line profile levels. Compared with five STF methods, STF-TH demonstrated significant improvements, ranging from 12% to 261.28% across six image quality evaluation metrics. In addition, STF-TH achieved superior spatial texture preservation and temporal color transformation, with improvements of 51.85% and 59.07%, respectively. Furthermore, STF-TH significantly improved the subsequent classification accuracy, with the F1-score and the overall accuracy improved to 89.88% and 93.87%, respectively. Notably, these STF-based improvements in STF-TH incurred negligible additional time consumption. Experimental results confirm that STF-TH is an efficient and effective model for temporal harmonization, considering potential problems of noise, patch effects, and spatial resolution degradation in traditional STF processing. STF-TH is expected to be applied to large-scale high-resolution annual land-cover monitoring.