Soil organic carbon (SOC) is a crucial indicator for maintaining soil fertility and regulating carbon balance in black soil regions. However, its strong spatial heterogeneity and the limited capacity of remote sensing feature extraction often lead to systematic mapping errors, typically manifested as the underestimation of high values and overestimation of low values. To address this issue, we propose an SOC mapping framework that integrates prior geographic knowledge with deep learning, and develop a GMM-AG-CNNLSTM model incorporating fuzzy clustering and spatiotemporal feature extraction. The framework was applied to typical black soil regions in Northeast China and North America. A total of 2,616 surface SOC samples (0–20 cm) were compiled to build a multi-source spatiotemporal feature set. The approach first employs a Gaussian mixture model (GMM) to partition SOC levels and mitigate prediction bias caused by spatial heterogeneity. Subsequently, a weighted attention mechanism, convolutional neural networks (CNN), and long short-term memory (LSTM) networks are combined to achieve deep spatiotemporal feature fusion and generate SOC distribution maps at a 30 m resolution. Results demonstrate that the GMM-AG-CNNLSTM model achieved prediction accuracies of R2 = 0.73/RMSE = 5.42 g/kg in Northeast China and R2 = 0.70/RMSE = 5.89 g/kg in North America, outperforming random forest and conventional deep learning models, with greater stability in both high- and low-SOC regions. Spatial analysis further revealed that SOC in Northeast China exhibited higher mean values and a larger proportion of high-value areas compared with North America, though with a wider distribution of low-value areas. This study presents a high-accuracy SOC remote sensing mapping method that can provide valuable support for carbon sequestration assessment and degradation monitoring in black soil regions.