On blending Landsat-MODIS surface reflectances in two landscapes with contrasting spectral, spatial and temporal dynamics

[摘要] Single-sensor monitoring systems are constrained by their inherent data characteristics in the spectral, radiometric, spatial and temporal domains.To better capture effective domain dynamics, a suite of algorithms have recently been developed that blend high spatial resolution / low temporal density imagery (e.g., Landsat TM/ETM+) with low spatial resolution / high temporal density imagery (e.g., MODIS).The output is a series of simulated surface reflectance imagery with high spatial resolution and high temporal density.As yet, there has been no definitive comparison of advanced algorithms (i.e., STARFM and ESTARFM) to each other over sites with vastly different domain characteristics, or when compared to computationally efficient simple algorithms used as a performance benchmark.Using a series of Landsat-MODIS pairs, we evaluated STARFM and ESTARFM against two simple algorithms to characterise their abilities to simulate Landsat-like surface reflectances.This was performed at two landscapes with contrasting spectral, spatial and temporal dynamics: (i) Coleambally Irrigation Area is a moderately temporally dynamic, more spatially heterogeneous site; and (ii) Lower Gwydir Catchment is a very temporally dynamic, less spatially heterogeneous site due to a large flood event.The blending algorithms were evaluated: (i) over entire sites; (ii) for subsets of hydrologically active areas within both sites (summer irrigated crops in Coleambally; and flooding / inundation in Gwydir); and (iii) for nine specific land cover types in Coleambally only.Importantly, for the whole Coleambally, the simple algorithms were as accurate as the advanced algorithms, however when assessing the summer irrigated crops only, ESTARFM provided optimal results.For the nine land cover types at Coleambally, mixed results were obtained as influenced by the spectral, spatial and temporal properties of the specific land cover.Over the entire Gwydir and for the highly dynamic flooded area, STARFM was the best algorithm, especially in the longer wavelengths.We found that ESTARFM was not always more accurate than STARFM, and that the advanced algorithms were not always more accurate than simple algorithms.Our study highlighted the need for: (i) development of a single metric to summarise the spectral-spatial-temporal variance of input imagery; and (ii) a suite of standardised approaches to assess the accuracy of blending algorithms.Successfully addressing both topics will aid the blending algorithm selection process, as our results clearly demonstrated that no single blending algorithm was optimal under all conditions tested.

[发布日期] 2012-08-31 [发布机构] CSIRO

[学科分类] 地球科学(综合)