Block‐Wise Data Synthesis to Achieve 1 Km Resolution Monthly Rainfall Forecast for Mainland China With Small Computing Resources

ABSTRACT

This study proposes a block‐wise data synthesis approach for operational high‐resolution rainfall prediction in resource‐constrained environments. To address the infeasibility of conventional full‐resolution LSTM training under limited computational capacity, the proposed method decomposes raw data into multiple subblocks for independent prediction, followed by reassembly and reconstruction to achieve high‐resolution output. Monthly precipitation data from the National Tibetan Plateau Data Center were employed alongside Long Short‐Term Memory (LSTM) networks for model development. During the data preprocessing stage, average pooling was applied to reduce the spatial resolution of the training data to one‐fifth of the original to accommodate available hardware constraints. The core model training was conducted on the compressed block‐wise data, with spatial inversion subsequently applied to restore high‐resolution outputs, thereby establishing a complete workflow encompassing “data decomposition–block‐wise training–reassembly and reconstruction.” Experimental results demonstrate that the proposed method achieves acceptable prediction accuracy for monthly rainfall forecasting tasks while significantly reducing per‐GPU memory consumption and training time. Compared with conventional full‐resolution training approaches, the method enables high‐resolution modeling over larger spatial extents under equivalent hardware conditions. The experimental results validate the engineering feasibility and operational deployment value of the block‐wise data synthesis approach in resource‐constrained scenarios. Future research will explore the incorporation of additional meteorological variables and model coupling strategies to further enhance the capability of this methodology. Furthermore, monthly rainfall forecasting is directly relevant to engineering practices such as water resources management, flood prevention and disaster mitigation, and agricultural production. The forecasting workflow presented in this study can support reservoir system operation at the basin scale, urban flood control infrastructure design, and optimization of disaster preparedness plans, underscoring its practical value within operational systems.