Machine Learning Techniques for Downscaling SMOS Satellite Soil Moisture Using MODIS Land Surface Temperature for Hydrological Application

Many hydrologic phenomena and applications such as drought, flood, irrigation management and scheduling needs high resolution satellite soil moisture data at a local/regional scale. Downscaling is a very important process to convert a coarse domain satellite data to a finer spatial resolution. Three artificial intelligence techniques along with the generalized linear model (GLM) are used to improve the spatial resolution of Soil Moisture and Ocean Salinity (SMOS) derived soil moisture, which is currently available at a very coarse scale of similar to 40 Km. Artificial neural network (ANN), support vector machine, relevance vector machine and generalized linear models are chosen for this study to integrate the Moderate Resolution Imaging Spectroradiometer (MODIS) Land Surface Temperature (LST) with the SMOS derived soil moisture. Soil moisture deficit (SMD) derived from a hydrological model called PDM (Probability Distribution Model) is used for the downscaling performance evaluation. The statistical evaluation has also been made with the day-time and night-time MODIS LST differences with the mean day and night-time PDM SMD data for the selection of effective MODIS products. The accuracy and robustness of all the downscaling algorithms are discussed in terms of their assumptions and applicability. The statistical performance indices such as R (2) , %Bias and RMSE indicates that the ANN (R (2) = 0.751, %Bias = -0.628 and RMSE = 0.011), RVM (R (2) = 0.691, %Bias = 1.009 and RMSE = 0.013), SVM (R (2) = 0.698, %Bias = 2.370 and RMSE = 0.013) and GLM (R (2) = 0.698, %Bias = 1.009 and RMSE = 0.013) algorithms on the whole are relatively more skillful to downscale the variability of the soil moisture in comparison to the non-downscaled data (R (2) = 0.418 and RMSE = 0.017) with the outperformance of ANN algorithm. The other attempts related to growing and non-growing seasons have been used in this study to reveal that season based downscaling is even better than continuous time series with fairly high performance statistics.