Mehdi Mahbod; Saeedeh Safari; Mohammaf Rafie Rafiee
Abstract
Determining precipitation spatial pattern in a catchment is necessary for the calculation of hydrologic quantities such as runoff flow and soil moisture content. Sparse meteorological stations as well as spatial variability of precipitation are major obstacles for accurate spatial estimation of precipitation. ...
Read More
Determining precipitation spatial pattern in a catchment is necessary for the calculation of hydrologic quantities such as runoff flow and soil moisture content. Sparse meteorological stations as well as spatial variability of precipitation are major obstacles for accurate spatial estimation of precipitation. The development of remote sensing technology and the possibility of using satellite precipitation products has facilitated attaining spatial precipitation patterns. However, low spatial resolution of satellite precipitation products highlights the need for downscaling methods. Nineteen predictive models were fitted using Regression Learner toolbox in MATLAB software. Annual TRMM precipitation data were downscaled from 2001 to 2017 using Normalized Difference Vegetation Index (NDVI), land surface temperature, land elevation and coordination. Models are divided into five general categories: Linear Regressions, Decision Trees, Support Vector Machines, Ensemble models and Gaussian Process Regression models. Comparing the downscaled TRMM data with gauges data, Boosted Ensemble model had the lowest root mean square error and highest correlation coefficient. On the other hand, two methods of Geographical Distance Adjustment (GDA) and Geographical Ratio Adjustment were compared for calibrating the downscaled precipitation. Smaller errors were obtained using GDA in all models.
