Enhanced Landsat surface reflectance prediction considering land cover change by using an ensemble of spectro-temporal and spectro-spatial predictions

Abstract

The cloud-induced contaminations limit the availability of Earth observation data acquired by Landsat satellites. In the field of cloud-contaminated image prediction and restoration, abrupt land cover change is an important but usually unconsidered issue. Based on the existing Time Series Models and Continuous Change Detection algorithm (TSM-CCD), an enhanced TSM-CCD algorithm (eTSM-CCD) is proposed in this paper to improve Landsat surface reflectance (SR) predictions based on all available Landsat data. The proposed eTSM-CCD method integrates time-series model estimation in the temporal domain and similar pixel replacement in the spatial domain to both address the issue of abrupt land cover changes and improve the prediction accuracy of SR. The proposed method was applied to Landsat time-series data in two study areas in the United States. The results indicated that the proposed method achieved higher overall and band-wise (especially NIR band) prediction accuracies than the original TSM-CCD algorithm. Specifically, the accuracies were improved more significantly for land cover types with distinct phenological characteristics at the peak of the growing season. In addition, compared to classical spatiotemporal prediction algorithms, such as modified neighborhood similar pixel interpolator (MNSPI) and enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM), the proposed method also yielded higher prediction accuracies and performed much better in predicting abrupt changes in land cover. The eTSM-CCD algorithm proposed in this paper is promising for the prediction and restoration of surface reflectance, especially in the vegetated areas during the peak of the growing season or in areas undergoing abrupt land cover changes.