Spatiotemporal Image Fusion With Spectrally Preserved Pre-Prediction: Tackling Complex Land-Cover Changes

Spatiotemporal image fusion enables the generation of time-series high spatial resolution (HR) images for monitoring fine-scale land surface dynamics, particularly for long-term (including historical) changes. Despite remarkable improvements, current spatiotemporal fusion methods still face challenges in accurately predicting complex land-cover changes. This article proposes a spatiotemporal image fusion model enhanced by spectrally preserved pre-prediction (PreSTFM) to improve the accuracy of land-cover change detection and reconstruction. The temporary pre-predicted HR image achieves a high level of spectral fidelity, specifically for land-cover changes, by introducing a multiband spectral mapping approach. Moreover, the pre-prediction plays a crucial role in establishing land-cover change-based constraint conditions to address the issue of incorrect similar pixels found in weighting-based methods. In addition to land-cover changes, PreSTFM can ensure that predicted changes align more accurately with actual changes (also including phenological changes) occurring in the landscape owing to spectrally preserved pre-prediction and spatial filtering mechanisms. The proposed PreSTFM was tested using three time-series Landsat and Moderate-Resolution Imaging Spectroradiometer (MODIS) datasets, compared with a flexible spatiotemporal data fusion (FSDAF) model and a robust adaptive spatial and temporal fusion model (RASTFM). The results indicate that PreSTFM outperforms FSDAF and RASTFM, yielding a root-mean-square error (RMSE) reduction of 6.1%-22.7% and 10.4%-27.5%, respectively. In addition, the PreSTFM predictions visually illustrate marked enhancements in capturing complex land-cover changes. These promising improvements highlight an effective and robust way of treating land surface changes, especially land-cover changes in spatiotemporal image fusion.