K. N. Toosi University of Technology Numerical Methods in Civil Engineering 2345-4296 10 4 2026 06 01 A Spatiotemporal CNN–LSTM Framework for Daily Precipitation Bias Correction Using Satellite and Reanalysis Data 1 11 250652 10.61882/NMCE.2601.1124 EN Seyed Mohammad Miri Ph.D. Candidate, Faculty of Civil Engineering, K.N. Toosi University of Technology, Tehran, Iran Mohammad Reza Kavianpour Professor, Faculty of Civil Engineering, K.N. Toosi University of Technology, Tehran, Iran Journal Article 2026 01 27 Accurate daily precipitation estimates are critical for hydrological and civil engineering applications, especially in regions affected by short-duration rainfall events. Satellite-based products provide wide coverage but often show systematic biases compared with ground observations. This study presents a spatiotemporal convolutional neural network–long short-term memory (CNN–LSTM) framework for daily precipitation bias correction in Golestan Province, northern Iran, using satellite and reanalysis data. The proposed framework integrates PERSIANN (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks) satellite precipitation with ERA5 reanalysis precipitation and 2-m air temperature to improve station-scale rainfall estimation at several synoptic stations. In order to represent the spatial organization of precipitation systems vividly, gridded precipitation from neighboring upstream cells was incorporated into the model structure. This configuration accounts for the dominant west-to-east movement of rainfall systems across the region, where precipitation commonly arrives with temporal delay and reduced intensity. In addition, temporal variations in near-surface air temperature were included to provide supplementary physical information associated with rainfall development. Model performance is evaluated using root mean square error (RMSE), correlation coefficient (CC), and mean error (ME). Categorical skill is assessed using probability of detection (POD) and false alarm ratio (FAR). The results indicate that the CNN–LSTM model consistently outperformed both conventional bias-correction approaches and baseline deep learning models. Incorporating spatial information from adjacent grid cells reduces errors and improves temporal consistency. Across the evaluated stations, the proposed model achieved the lowest RMSE values, with error reductions reaching 23% relative to the DNN benchmark. The framework also demonstrated improved capability in capturing the spatiotemporal evolution of precipitation events. Overall, the proposed approach provides a transferable and operationally practical framework for enhancing daily precipitation estimates in data-scarce regions. The improved precipitation estimates can support more reliable hydrological simulations, flood assessment, and water resources planning. https://nmce.kntu.ac.ir/article_250652_97182be604de0627a01b76cbcff49f3b.pdf