Abstract: Accurate extraction of spatial information on precipi tation is crucial in improving the accuracy of runoff prediction. Taking the Ganjiang River basin as a case study, this paper develops five precipitation spatial information processing schemes (i.e., raw images, wavelet decomposition, statistical features, mean areal precipitation, and regional partitioning) based on a long short-term memory model. The aim is to investigate how different approaches to capturing spatial variability in precipitation affect the performance of LSTM-based runoff prediction. The results show that, compared with the direct use of raw images, integrated methods employing wavelet decomposition and statistical feature extraction improve the Nash–Sutcliffe efficiency by 11.5% and 17.9%, respectively, during the testing phase, while also enhancing model stability and interpretability. Furthermore, various regional partitioning strategies incorporating land use, soil type, and other underlying surface factors capture the spatial heterogeneity of precipitation response more effectively compared to approaches that use basin-wide or regional average values as inputs, demonstrating robust performance across different flow regimes and significantly improving the ability to capture both high- and low-flow conditions. The result show that incorporating precipitation spatial information into LSTM-based rainfall—runoff models can markedly enhance predictive accuracy.