A comparison of STARFM and an unmixing-based algorithm for Landsat and MODIS data fusion

The focus of the current study is to compare data fusion methods applied to sensors with medium- and high-spatial resolutions. Two documented methods are applied, the spatial and temporal adaptive reflectance fusion model (STARFM) and an unmixing-based method which proposes a Bayesian formulation to incorporate prior spectral information. Furthermore, the strengths of both algorithms are combined in a novel data fusion method: the Spatial and Temporal Reflectance Unmixing Model (STRUM). The potential of each method is demonstrated using simulation imagery and Landsat and MODIS imagery. The theoretical basis of the algorithms causes STARFM and STRUM to produce Landsat-like reflectances while preserving the spatial patterns found in Landsat images, and the unmixing-based method to produce MODIS-like reflectances. The ability of fused images to capture phenological variations is also assessed using temporal NDVI profiles. Temporal profiles of STARFM NDVI closely resembled Landsat NDVI profiles. However, the unmixing-based method and STRUM produced a more accurate reconstruction of the NDVI trajectory in experiments simulating situations where few input high-resolution images are available. STRUM had the best performance as it produced surface reflectances which had the highest correlations to reference Landsat images. The results of this study indicate that STRUM is more suitable for data fusion applications requiring Landsat-like surface reflectances, such as gap-filling and cloud masking, especially in situations where few high-resolution images are available. Unmixing-based data fusion is recommended in situations which downscale the spectral characteristics of the medium-resolution input imagery and the STARFM method is recommended for constructing temporal profiles in applications containing many input high-resolution images.