High spatial and temporal resolution remote sensing data play an important role in monitoring the rapidly changing information regarding the earth’s surface. However,the spatial and temporal resolutions of a remote sensing image for a specific sensor contain irreconcilable conflict. Fusing the spatial and temporal features of different remote sensing images to generate high-spatialtemporal remote sensing data is an effective method to solve the contradiction. This paper aims to improve the fusion data performance of STARFM in heterogeneous areas by downscaling rather than directly resampling a coarse pixel.The combination of the downscaling mixed pixel method and the spatial and temporal adaptive reflectance fusion model（ STARFM）（ CDSTARFM） was proposed. In this approach,the downscaling mixed pixel method first reduces the MODIS data that participated in the fusion. Then,the downscaled MODIS data replaces the direct resample MODIS data that appeared in the original STARFM. Finally,the subsequent steps of the STARFM are completed to predict Landsat-like images. The proposed algorithm produced high-resolution temporal synthetic Landsat 8 data based on Landsat 8 and MODIS remote sensing images.Results show that the CDSTARFM was more accurate than STARFM and downscaling methods. The downscaled data used in the CDSTARFM can more fairly reflect surface information than the resampled data of MODIS,which increase the probability of"pure pixels"and allow the CDSTARFM method to more easily determine the"pure"similar pixels in the search window. Therefore,the three indicators [i. e.,correlation coefficient（ r）,root mean square error（ RMSE）,and the Erreur Relative Globale Adimensionalle de Synthèse（ ERGAS） ] as well as the scatterplots and visual effect of synthetic images for the CDSTARFM are better than those obtained by STARFM and downscaling methods. Moreover,the optimal window size（ 11 × 11 OLI pixels） of the CDSTARFM is smaller than that of the STARFM（ 31 × 31 OLI pixels）. The accuracy of the CDSATRFM at the same window size is higher than that of the STARFM. The predicted NIR band is evaluated as an example in this study. The correlation coefficients（ r） for the CDSTARFM,STARFM,and downscaling methods were 0. 96,0. 95,0. 90,respectively; RMSEs were 0. 0245,0. 0300,0. 0401,respectively; and ERGAS were 0. 5416,0. 6507,0. 8737,respectively. Moreover,synthetic images effectively eliminated the"homogeneous spot"that appeared in the fusion images predicted by the downscaling algorithm and the"boundary of MODIS pixel"that appeared in the synthetic images produced by the STARFM.The temporal-spatial fusion of remote sensing data is an effective approach to solve the conflict of temporal and spatial resolutions of a sensor. This paper used the CDSTARFM algorithm,which combines downscaling method and STARFM to fuse remote sensing data. The CDSTARFM algorithm was verified by the experimental data of Landsat 8 and MODIS images,and it was compared with the STARFM and downscaling methods. The conclusions are listed bellow:（ 1） CDSTARFM using the downscaled data presented improved results to replace the directly resampled data used in STARFM. The three indicators（ i. e.,r,RMSE,and ERGAS） and the scatterplots for the CDSTARFM are the best compared with those for STARFM and downscaling methods.（ 2） The window size at which the best synthetic images were predicted by CDSTARFM is smaller than that by STARFM. In addition,CDSTARFM exhibits better accuracy than STARFM and downscaling methods at the same window size.（ 3） The synthetic images produced by CDSTARFM are more visually similar to the reference images,especially in the fragmented regions. CDSTARFM can eliminate the homogeneous"plot"in the synthetic images that were generated by downscaling methods and the "MODIS pixel boundary"in the prediction images generated by STARFM in fragmented landscapes.

decomposition of mixed pixel;downscaling;STARFM;remote sensing;data fusion;CDSTARFM