Bias correcting the precipitation dynamics of regional climate models via kernel-aware 2D convolutional-long short-term memory

Some climate models face challenges covering the globe on a finer spatial scale. As a result, local studies are hindered by this limitation. This study introduces a novel spatial-based (kernel-aware) 2D Convolutional-Long Short-Term Memory (Conv-LSTM) network to enhance and bias correct spatial dynamics and generate precipitation products from Regional Climate Models (RCMs). The proposed network used 3 x 3 kernels, known as pixels that comprise nine grids for each specific point, which conduct convolutional layers to extract the features from the broad area (75 x 75 km), and LSTM networks for handling temporal dependencies. In this way, the RCM-based precipitation data were used as reference inputs, and gridded precipitation observation as target values. Since the precipitation products from the Coupled Model Intercomparison Project Phase 5 (CMIP5) of RCMs consisted of systematic biases, Empirical Quantile Mapping (EQM) was first used as the bias correction method as the pre-bias correction. This study applied 360 monthly observation precipitation and 460 bias- corrected RCM grid points covering Southern Alberta spanning from 1962 to 2006. Moreover, the proposed model was compared with the classical Feed Forward Neural Network (FFNN). Furthermore, the network's capability spanned to the future, using Representative Concentration Pathway 4.5 till the end of this century. The results demonstrated that the proposed novel network could capture adjacent precipitation impacts on the target point and produce observation-like products with more precision by the Root Mean Squared Error (RMSE) and Determination Coefficient (DC) of 17.65 mm, 17.07 mm, 14.74 mm, and 0.60, 0.71 and 0.85 for high, low, and normal precipitation conditions, respectively.