Abstract: Cloud Base Height (CBH) is a key parameter for characterizing the vertical structure of clouds and holds significant application value in areas such as precipitation forecasting, identification of suitable conditions for weather modification operations, and aviation safety. However, existing observation methods face challenges in simultaneously achieving spatial coverage, temporal resolution, and accuracy, making large-scale, all-weather, and high-precision CBH retrieval technically difficult. To address this, this paper proposes a deep learning-based multi-modal feature fusion method for retrieving cloud base height, named Deep-Cloud Base (DCB). The method uses multi-channel observations from the Advanced Geosynchronous Radiance Imager (AGRI) aboard the FY-4B (Fengyun-4B) satellite as the core input, integrates multi-source information including cloud masks and observation geometry, and constructs an end-to-end retrieval model. The model is supervised and trained using CBH values derived from ground-based millimeter-wave cloud radars as ground truth. Trained and optimized on over 1.5 million satellite–cloud radar matched samples from March 2024 to March 2025, the model achieves a Root Mean Square Error (RMSE) of 1959.7 m, a Mean Absolute Error (MAE) of 1138.4 m, with Pearson correlation coefficients of 0.83 on an independent test set. A scenario-based evaluation of the retrieval results shows that the method performs best under daytime conditions, for mid-level clouds, and in single-layer cloud scenarios, while exhibiting greater uncertainty in the presence of complex multi-layer cloud structures. The nationwide, high spatiotemporal resolution CBH product generated by the DCB algorithm can provide critical support for applications such as identifying weather modification operation conditions.