Assessing the accuracy of blending Landsat–MODIS surface reflectances in two landscapes with contrasting spatial and temporal dynamics: A framework for algorithm selection

Blending algorithms model land cover change by using highly resolved spatial data from one sensor and highly resolved temporal data from another. Because the data are not usually observed concurrently, unaccounted spatial and temporal variances cause error in blending algorithms, yet, to date, there has been no definitive assessment of algorithm performance against spatial and temporal variances. Our objectives were to: (i) evaluate the accuracy of two advanced blending algorithms (STARFM and ESTARFM) and two simple benchmarking algorithms in two landscapes with contrasting spatial and temporal variances; and (ii) synthesise the spatial and temporal conditions under which the algorithms performed best. Landsat-like images were simulated on 27 dates in total using the nearest temporal cloud-free Landsat–MODIS pairs to the simulation date, one before and one after. RMSD, bias, and r2 estimates between simulated and observed Landsat images were calculated, and overall variance of Landsat and MODIS datasets were partitioned into spatial and temporal components. Assessment was performed over the whole study site, and for specific land covers. Results addressing objective (i) were that: ESTARFM did not always produce lower errors than STARFM; STARFM and ESTARFM did not always produce lower errors than simple benchmarking algorithms; and land cover spatial and temporal variances were strongly associated with algorithm performance. Results addressing objective (ii) indicated ESTARFM was superior where/when spatial variance was dominant; and STARFM was superior where/when temporal variance was dominant. We proposed a framework for selecting blending algorithms based on partitioning variance into the spatial and temporal components and suggested that comparing Landsat and MODIS spatial and temporal variances was a practical method to determine if, and when, MODIS could add value for blending.

► STARFM, ESTARFM and two simple benchmarking blending algorithms were evaluated.
► Blending algorithm performance was related to spatial and temporal variances.
► ESTARFM and STARFM accounted for spatial and temporal variance better, respectively.
► STARFM sometimes outperformed ESTARFM; simple algorithms were sometimes most accurate.
► Comparing Landsat and MODIS spatial and temporal variances aids algorithm selection.