Abstract: To improve the spatial resolution of satellite precipitation products to meet the need for fine hydrological research, a convolutional neural network-based precipitation downscaling model is developed. It adopts the relationships of precipitation versus NDVI, DEM, slope, slope direction, latitude, and longitude from the Global Precipitation Observation Program multi-satellite precipitation product (IMERG) of better accuracy for regions in China. We also examine its overall performance and variations in its parameters for different timescales of a year, season, month, and ten-day. Results show that agreement indices of annual, seasonal and monthly downscaled results exceed 0.94, 0.89, and 0.69 respectively in comparison with original data; the model can also characterize precipitation at the ten-day scale. Compared with China Gauge-Based Daily Precipitation Analysis (CGDPA) data, the average agreement indices of annual, seasonal, monthly and ten-day downscaled results are 0.58, 0.78, 0.68, and 0.47 respectively; the similarity of model parameters increases gradually with the depth of the model layers. This study shows that this deep learning model has good convergence and good potential for basin-wide downscaling applications.

Key words: IMERG, CGDPA, convolutional neural networks, Luanhe River basin, statistic downscaling