Blending and Downscaling of Landsat and MODIS Surface Reflectance for Water Body Delineation: A Comparison of Index-Simulate and Simulate-Index Methods

Single-sensor satellite remotely sensed data are typically either high temporal and low spatial resolution (e.g. MODIS) or low temporal and high spatial resolution (e.g. Landsat). Blending algorithms have been developed to overcome this limitation by predicting the composition of higher spatial and temporal resolution band data by blending individual corresponding bands of two sensors. The objective of this paper was to evaluate the accuracy of two advanced algorithms (STARFM and ESTARFM) in blending single bands and indices to downscale MODIS pixels (250-500 m) to Landsat resolution (28.5 m). We test two approaches to predicting indices: i) Index-Simulate (IS)-(i.e., indices directly predicted including EVI, NDVI NDWI24 and NDWI 27) from Landsat-MODIS pairs) and ii) Simulate- Index (SI)-(i.e. indices were calculated from five predicted bands (Blue, Green, Red, Near-Infrared and Mid-Infrared). Landsat-like images and indices (IS and SI) were predicted for 18 dates by using 20 pairs of cloud-free Landsat 5 TM and Aqua MODIS images and compared with observed Landsat images and indices. Based on RMSD (accuracy of predicted and observed bands and indices), pixel-to-pixel accuracy of each prediction and R-squared differences of predicted and observed pixels, ESTARFM produced a lower error than STARFM in predicting all four tested indices. Results of IS and SI methods showed that, both algorithms predict indices in IS method with higher accuracy than using SI method. That is, if interested in using indices in applications it best to calculate the index at the two resolutions then use the algorithms to simulate the index as opposed to simulating the individual bands then subsequently calculating the index. This study shows that, the high spatio-temporal resolution predicted water index (NDWI) can be used in water resources applications. Landsat-like daily water indices simulated by using blending Landsat and MODIS data provided daily flood inundation footprint. These outputs provide new insight into the application of remote sensing data in conjunction with hydrodynamic models for better understanding of hydrological processes in data sparse environments such as central Australia with complex anabranching river system.