On the spatial and temporal perfomance of ESTARFM downscaling method for generating Landsat-like imagery

Satellite time series data play a key role in characterizing land surface change and monitoring of short and long-term land cover change processes over time. While coarse spatial resolution optical sensors (e.g. MODIS) can provide appropriate time series data, the temporal resolution of high to intermediate spatial resolutions sensors (1-100 m e.g. Landsat) does not allow for having temporally frequent measurements because of the orbital configuration of such sensors and cloud contamination. A promising approach for addressing this challenge and producing Landsat-like imageries is the blending of data from coarse spatial resolution sensors like MODIS. Among different approaches proposed in the literature, the ESTARFM model has been reported to outperform other models in generating Landsat-like imageries with reasonable accuracy over heterogeneous areas. Despite the large body of studies implementing ESTRAFM for downscaling MODIS data, quantitative evaluation of the model under different conditions has not yet been investigated. This study quantitatively evaluates model performance over different land cover types, resampling methods and time-difference analysis between input and synthetic images. The results demonstrated that employing bilinear resampling in the ESTARFM produces results slightly better than nearest neighbor and cubic resampling methods. Moreover, the ESTARFM model accurately predicts Landsat-like surface reflectance images with RMSE better than 0.02 and correlation more than 90% over different land cover types. However, the model performance significantly degrades as the time difference between the input and synthetic images increases.