Bias Correction of Global Blended Satellite Precipitation Products over Taiwan with Deep Learning

Although rain gauge networks can offer highly accurate surface precipitation estimates but their global distribution is sparse and nonuniform from nation to nation which is not able to show the whole picture of precipitation dynamics of the earth, thereby creating discontinuous observations. Typically, weather radars compensate for the range of ground gauge coverage with high spatial and temporal resolution, however, exhibiting their poor performance over space having extreme topography such as mountain areas. More importantly, neither rain gauge networks nor weather radars can acquire precipitation observations over remote oceans, especially in tropical areas where the most precipitation happen but contribute to the significant quantity of the planet water cycle. By contrast, spaceborne sensors for precipitation estimating that leverage the extensive scan footprint collaborating with orbital satellites have been extremely successful for establishing a seamless precipitation map for weather prediction and climate monitoring at both local and global scales.

However, inherent imperfections in the passive microwave (PMW) sensors, coupled with retrieval algorithms, interpolation processes and data augmentation techniques (level 2 to 3) such as uncertainties in could-motion propagation, can introduce substantial errors in satellite precipitation products (SPPs). In addition, the presence of strong interactions between lands and oceans can exacerbate the uncertainty associated with SPPs. Several well-known phenomena exist in SPPs including but not limited to precipitation intensity overestimation/underestimation during shallow/heavy precipitation events, respectively, and tending to underestimate precipitation in mountainous regions, resulting in positive and negative biases. Although currently operational global SPPs have integrated various bias-correction postprocessing techniques such as rain gauge correction, PDF matching and climatological calibration, the adjustments are at coarse scales and the improvements are severely restricted in local areas. There hence is an urgent need to bias-correct these erroneous estimates at higher resolutions with acceptable tolerance locally.

To overcome the limitations of traditional methods, we propose a deep learning-based approach that utilizes a deep convolutional neural network (CNN) as a spatial precipitation pattern extractor to regress error structures existing in the NOAA Climate Prediction Center (CPC) morphing technique (CMORPH) and the NASA Global Precipitation Measurement (GPM) Integrated Multi-satellite Retrievals for GPM (IMERG), to the high-resolution ground-based reference data produced from the operational Quantitative Precipitation Estimation and Segregation Using Multiple Sensors (QPESUMS) system of Taiwan to perform bias-correction at daily scale. The entire framework accounts for Taiwan’s seasonal precipitation patterns and topography features with customized loss functions and training techniques, to deliver a stable bias-correction solution. This framework is the successor to our precious research in Northern California.