Comparison of multiple linear regression and artificial neural network models for downscaling TRMM precipitation products using MODIS data

Precipitation plays a significant role to energy exchange and material circulation in Earth’s surface system. According to numerous studies, traditional point measurements based on rain gauge stations are unable to reflect the spatial variation of precipitation effectively. On the other hand, satellite remote sensing could solve this limitation by directly providing spatial distribution of rainfall over large areas. During the last years, the Tropical Rainfall Measuring Mission (TRMM) has provided researchers with a large volume of rainfall data used for the validation of atmospheric and climate models. However, due to its coarse resolution (0.25°) the improvement of its resolution appears as a fundamental task. The main aim of this study is to compare two different integrated downscaling-calibration approaches namely multiple linear regression analysis and artificial neural networks for downscaling TRMM 3B42 precipitation data. The statistical relationship among TRMM precipitation data and different environmental parameters such as vegetation, albedo, drought index and topography were tested in the island of Crete, Greece. Free distributed satellite data of coarse resolution such as those of MODIS sensor were incorporated in the overall analysis. Multiple linear regression as well as artificial neural network models was developed and applied, and extensive statistical analysis was performed by downscaling the TRMM products. The downscaled precipitation estimates as well as the TRMM products were subsequently validated for their accuracy by using an independent precipitation dataset from a ground rain gauge network. The downscaling procedure succeeded to significant improvements of monthly precipitation estimation (100 % improvement in terms of spatial resolution) in terms of spatial analysis with means of satellite remote sensing.