Spatial downscaling of the GCMs precipitation product over various regions of Iran: Application of Long Short-Term Memory model

Global climate models (GCMs) are important tools for obtaining past and future climate information from sophisticated mathematical and physical equations. Despite their extensive development, these models still exhibit substantial uncertainty due to their large scale. It is critical to select the right downscaling methods for yielding accurate and reliable climate change projections and properly analyzing the impact of climate change on regional and local scales. This study analyzes the applicability of the Long Short-Term Memory (LSTM) artificial neural network for downscaling precipitation in ten phase-6 climate models across various climates in Iran. This study also analyzes the performance of GCMs across different climates with 24 Iranian synoptic stations as target data from 1987 to 2014. Ten models were utilized for this purpose: GCM, CESM2-WACCM, FIO-ESM-2-0, GFDLESM4, INM-CM5-0, IPSL-CM6A-LR, MPI-ESM1-HR, and MRI-ESM2-0. The CESM2 and FIO-ESM-2-0 models demonstrate better outputs in mountainous regions in such a way that the correlation of those models are 0.43 and 0.36 respectively. The LSTM significantly enhanced the production by over 150% to the range of 0.45-0.58. The mean of the 10 GCMs in the eastern areas (G3-G5-G7) showed a satisfactory value of 0.62, with their enhancement rates in these areas being less than those in other regions. The Caspian Sea regions (G6-G8) exhibited the highest MAE and RMSE indices with 146 and 99 respectively, with certain improvements after the neural model was implemented. The most significant improvement in these two indices is seen in the northwestern and western regions (G3, G5) with 0.62, 0.18, 0.2 in CC, RMSE, and MAE respectively, and certain parts of central Iran. The G1, G3, and G5 regions had the lowest values of MAE and RMSE errors. Overall, the LSTM model demonstrated excellent performance in downscaling, effectively improving correlation and minimizing errors.