Comparison of MODIS to Landsat-8 data Downscaling Algorithms for Evapotranspiration Estimation

Due to technical and financial limitations, it is not possible to simultaneously provide high spatial and temporal resolution by a sensor. There is always a trade-off between the spatial and temporal resolution of the sensors. For studies such as estimating evapotranspiration, land surface temperature with high temporal and spatial resolution is required; however, estimating actual evapotranspiration with high temporal and spatial resolution by a single sensor is not possible. Since high spatial and temporal resolution together increase the reliability of analyzing and extracting information from the image, so the best way to overcome this problem is to downscale images to high temporal and spatial resolutions. Downscaling is the process of converting images with low spatial resolution to images with high spatial resolution. So far, several methods have been proposed for downscaling. These methods differ for downscaling of the reflectance and thermal bands. Many studies that have been conducted so far on the actual evapotranspiration estimation, indicate the efficiency of SEBAL algorithm for this purpose. Therefore, in this study, in order to calculate the actual evapotranspiration, the SEBAL model was used and the products of different downscaling methods were given as input to this model. Assessing the accuracy of actual evapotranspiration ​​calculated using remote sensing data indicates the efficiency of products obtained from different methods. According to the studies conducted in this field, so far no study has been done on the combination of downscaled bands obtained from different downscaling methods applied on thermal data and non-thermal data in order to calculate the actual evapotranspiration. In this study, STARFM, ESTARFM and Regression algorithms were used to downscale the reflectance bands and SADFAT, Regression and Cokriging algorithms were used to downscale the thermal bands. Then the accuracy of the results was evaluated.

The study area is Amirkabir agro-industry located in the south of Khuzestan province, one of the seven companies for the development of sugarcane cultivation and ancillary industries (longitude 48.287100, and latitude 31.029696 degrees). The gross land area of this agro-industry is 15000 hectares and its net area is 12000 hectares which is divided into several 25-hectare plots. In this research, the images of MODIS located on Terra satellite and the images of OLI and TIRS sensors of Landsat 8 satellite were used. It is worth noting that the Landsat image for time 2 was used to evaluate the simulation results. The downscaling algorithms used in this research included STARFM, ESTARFM, and REGRESSION algorithms were applied on reflectance bands and SADFAT, Regression and Cokriging algorithms were used for thermal band downscaling. In order to conduct this research, first, various downscaling methods were applied on MODIS images to be downscaled to the images with Landsat spatial resolution. Then, using MODIS downscaled images, evapotranspiration values were calculated for different combinations of downscaled data using SEBAL method and the results were compared and evaluated with evapotranspiration obtained from Landsat images acquired at the same date as MODIS data.

In order to evaluate the results, the downscaled bands were visually and quantitatively compared with the corresponding bands of the Landsat image acquired on the same date. In order to compare these data quantitatively, the root mean square error (RMSE) and the coefficient of determination (R2) were used. According to the RMSEs, it can be concluded that the STARFM, ESTARFM, Regression, SADFAT and Cokriging downscaling algorithms all perform well. Among the methods applied to the reflectance bands, STARFM with the RMSE of 0.0180 had the best performance, followed by ESTARFM with the RMSE of 0.0186 and Regression with the RMSE of 0.0479. Among the methods applied to thermal bands, the SADFAT algorithm with the RMSE of 0.0224 had the best performance, followed by Cokriging with the RMSE of 0.0234 and Regression with the RMSE of 0.0464. It should be noted that the difference in outputs is very small, and given that the study area of ​​this study is a homogeneous area of ​​agricultural land cover including a single sugarcane crop. This issue can be the main reason for the close performance of downscaling methods and the high accuracy of their outputs. Moreover, according to the results obtained for evapotranspiration, ESTARFM / Regression, ESTARFM / SADFAT, STARFM / Regression and STARFM / SADFAT had the best performance with the lowest difference and the Regression / Cokriging method had the weakest performance, respectively.

This study can be concluded as follows:All downscaling algorithms used in this research had an acceptable performance in simulating Landsat bands.Among the reflectance band-related downscaling methods, STARFM had the best performance, followed by ESTARFM and Regression, respectively.Among the thermal band-related downscaling methods, the SADFAT algorithm performed best, followed by Cokriging and Regression.The use of STARFM algorithm for reflectance bands and SADFAT algorithm for thermal bands in homogeneous areas is recommended.The difference between the different combinations of methods for estimating actual evapotranspiration is small.